Keywords: Computer Vision for Manufacturing, Machine learning
Abstract: Teacher–student distillation based on normal samples has been widely adopted in Unsupervised Anomaly Detection (UAD). Incorporating Vision Transformers (ViTs) into this task enhances global representation capability; however, existing methods still suffer from feature instability caused by pseudo-activation phenomena and the high optimization cost of full-parameter fine-tuning. To address these issues, this paper proposes a Dual-Frozen ViT Distillation (DFVD) framework, which freezes both teacher and student backbones. The teacher branch introduces a test-time register mechanism to suppress pseudo-activations and provide stable targets, while the student branch integrates learnable prompt tokens and lightweight adapters for feature alignment. By optimizing only lightweight modules, the framework enables parameter-efficient adaptation while preserving the representation capability of pretrained ViTs. Experiments on the MVTec AD benchmark demonstrate the effectiveness of the proposed method for anomaly detection and localization.

Keywords: Computer Vision in Automation, Computer Vision for Manufacturing, Deep Learning in Robotics and Automation
Abstract: Monocular 6D pose estimation remains challenging due to low-quality 3D coordinate maps caused by occlusion, texture-less surfaces, and spatial detail loss in encoder-decoder networks. This paper presents an efficient monocular framework that improves pose accuracy by enhancing the quality of dense 3D coordinate maps via feature fusion and deformable decoupled regression. We propose a lightweight feature fusion module composed of long-range-aware feature fusion and multi-scale feature fusion, which integrates spatial details and long-range dependencies using deformable convolutions. Meanwhile, a disentangled pose regression network is designed to separately estimate rotation and translation for better optimization. Extensive experiments on Linemod and Linemod-Occluded datasets show that our method achieves competitive performance with state-of-the-art dense correspondence methods while maintaining real-time speed. Ablation studies verify the effectiveness of each proposed component, and the feature fusion module generalizes favorably to other representative pose solvers. Finally, our method provides an efficient and robust solution for monocular 6D pose estimation, with broad application prospects in real-world robotic scenarios.

Keywords: Computer Vision in Automation, Computer Vision for Manufacturing, Deep Learning in Robotics and Automation
Abstract: Real-time classification of welding defects is critical for automated quality monitoring in smart manufacturing, yet industrial datasets are typically small and severely imbalanced, making deep learning approaches prone to overfitting. This paper presents a visual classification pipeline for three weld quality categories: good welding, undercut, and lack of fusion. Aggressive augmentation and focal loss are employed to handle data limitations, alongside a systematic comparison of three lightweight transfer-learning architectures: MobileNetV3-Small, EfficientNet-B0, and ResNet18. EfficientNet-B0 achieves the best accuracy-efficiency trade-off, reaching 99.59% test accuracy with a macro F1 score of 0.996 and good classification of the minority undercut class, establishing it as the recommended architecture for practical industrial deployment. An original welding image dataset has been released on Kaggle to support reproducible research.

Keywords: Data fusion, Probability and Statistical Methods, Machine learning
Abstract: Milk quality inspection often involves multiple physicochemical and microbiological metrics, while measuring all metrics for every sample may significantly increase inspection cost and delay decision making. This paper presents a cost-aware sequential inspection framework for milk quality control based on deep reinforcement learning. Instead of treating quality assessment as a one-shot classification task, the proposed method formulates the inspection process as a sequential decision problem, where an agent dynamically selects the next metric to measure or stops early to issue a final decision. The final diagnosis is determined by domain rules defined on quality thresholds, while the reinforcement learning agent optimizes the sequence and inspection amount. Specifically, a Deep Q-Network (DQN) is employed to learn the inspection policy from historical data. The reward design jointly considers measurement cost, redundant metric selection, repeated measurements, and asymmetric misclassification risks, with a significantly larger penalty assigned to missed unqualified samples. In addition, a correlation-based redundancy penalty is introduced to encourage diverse metric selection, and failure-oriented sampling is adopted during training to alleviate class imbalance.

Keywords: Deep Learning in Robotics and Automation
Abstract: Imitation learning enables robots to acquire com plex manipulation skills from demonstrations, but its effec tiveness is limited by the cost of collecting high-quality data. Trajectory-level data augmentation methods alleviate this chal lenge by recombining expert demonstrations under varied ini tial states. However, such methods typically insert interpolations or other non-expert transition segments between disjoint parts, and such non-expert segments could reduce the quality of the generated data. This paper introduces Minimizing Inter polation (MinInter), an effective trajectory selection method that, for each sampled initial configuration, chooses the source demonstration requiring the least interpolation to form a com plete trajectory. By explicitly minimizing interpolations during data generation, MinInter produces higher-quality synthetic demonstrations while remaining compatible with existing data generation frameworks. Experiments on 12 manipulation tasks with 26 variants from the MimicGen benchmark show that MinInter consistently improves both data generation success rates and policy success rates, with the largest gains on contact rich, long-horizon and high-variance settings. Compared to the recent SkillGen framework, MinInter achieves higher policy success rates despite its conceptual simplicity, underscoring the value of interpolation minimization for data augmentation.

Keywords: Deep Learning in Robotics and Automation, Probability and Statistical Methods, Robust Manufacturing
Abstract: Single-pass uncertainty estimation for regression tasks is a critical component for the reliable deployment of Deep Neural Networks (DNNs) on robots. In industrial robotics and automation, where latency and power constraints are stringent, the ability to quantify model uncertainty in a single forward pass is essential for safe real-time decision-making. However, current single-pass uncertainty methods often struggle with outliers in the training data. In this work, we introduce a robust single-pass uncertainty framework that is specifically designed for resource-constrained settings. Our approach offers two key advantages: robust training through automatic outlier detection and handling, and efficient deployment via reliable single-pass uncertainty without the computational cost of multi-pass methods. To demonstrate the practical benefits of this method, we conducted robot experiments specifically designed to test performance under adversarial attacks and out-of-distribution conditions. Our results show that the proposed loss function allows the robot to maintain reliability and awareness of its own limitations, even when trained on data containing outliers and when faced with previously unseen data. By combining robust fitting with efficient estimation, our work enables a more dependable path for deploying DNNs on resource-constrained hardware.

Keywords: Deep Learning in Robotics and Automation, Reinforcement, AI-Based Methods
Abstract: As Fifth Generation (5G) networks evolve toward Sixth Generation (6G), base station energy consumption poses a critical challenge for sustainable wireless infrastructure. Traditional reinforcement learning approaches lack scalability, while centralized training exposes sensitive operational data in multi-operator deployments. To address these issues, we propose the Strategy-Aware Federated Split Decision Transformer (SA-FSDT) for privacy-preserving energy optimization in heterogeneous cellular networks. Our framework overcomes the spatial coordination limitations of existing methods by introducing a strategy-aware cross-attention mechanism. Specifically, small base stations compute local embeddings that are transmitted to a macro base station, which utilizes a hierarchical attention strategy to dynamically balance global network coordination with local capacity optimization. By transforming offline reinforcement learning into supervised sequence modeling, SA-FSDT eliminates risky online exploration while safeguarding proprietary information across operators. Extensive simulations demonstrate that SA-FSDT significantly reduces power consumption while simultaneously improving Quality of Service (QoS) satisfaction. Furthermore, it exhibits robust cross-density generalization without the need for retraining, validating its potential for collaborative, energy-efficient management in next-generation multi-operator networks.

Keywords: Deep Learning in Robotics and Automation, Reinforcement, Model Learning for Control
Abstract: Visuomotor reinforcement learning (RL) has shown a strong capability in enabling robotic agents to perceive their environments and generate actions to accomplish down-stream manipulation tasks. However, existing approaches often do shortcut learning, resulting in poor performance under out- of-distribution (OOD) conditions, while also inducing incomprehensible interpretations. To address these issues, we propose Concept-Guided Grounded Training (CGGT), a framework that leverages task-specific concepts from the environment to improve visuomotor learning, not only in terms of success rates but also in terms of generalization and interpretability. Specifically, our CGGT framework employs a single stage training scheme that integrates visual concept alignment at the feature extraction and spatial concept alignment at the decision-making network, along with the reward optimization. This approach enables robots to focus on action-relevant features and subsequently generate actions for task completion. Through our evaluations on eleven robotsuite tasks with multiple OOD scenarios in the RL-Vigen benchmark, CGGT outperforms state-of-the-art baselines in terms of generalization while also enabling failure inspection through concept visualizations. Finally, we also deploy our proposed method on the Franka Emika robot in four different tasks with real-world objects to test its feasibility in the real-world domain.

Keywords: Diagnosis and Prognostics, Machine learning, Failure Detection and Recovery
Abstract: While several classification strategies exist for bearing fault detection, degradation is a gradual process. This paper investigates whether switching models mid-degradation, guided by a measurable feature-separability metric, improves the sensitivity-latency trade-off for predictive maintenance on resource-constrained hardware. Using measurements which come from the NASA Intelligent Maintenance Systems (IMS) 2nd_test dataset (984 total files that measure Bearing 1, which has a fault on the outer race), three classifiers - Random Forest (RF), Logistic Regression (LR), and Support Vector Machine (SVM) - were evaluated with respect to their ability to detect faults at two stages of degradation. At the incipient stage, RF detected 27 out of 56 fault cycles (48.2% detection rate) while LR detected only 7 out of 56 (12.5%), representing a difference of 3.9x. At the advanced stage, all three classifiers converged to Area Under the Curve (AUC) = 1.000. Latency was measured using CPU cycle-counting directly from a real-world ESP32 microcontroller (Xtensa LX6, 240 MHz), not from simulation. LR had a latency of 184 ns while RF had a latency of 23.3 us, i.e., a 126x difference. SVM misclassified the fault on hardware at 96.5 us, exposing a single-precision numerical vulnerability near its decision boundary. Fisher Discriminant Ratio (FDR) analysis shows a 5.89x separability increase from incipient to advanced stage, and cross-bearing validation on IMS Bearing 2 yielded near-zero false-positive rates (RF: 1.8%; SVM/LR: 0.0%).

Keywords: Diagnosis and Prognostics, Manufacturing, Maintenance and Supply Chains
Abstract: Reducing downtime in manufacturing systems requires the rapid identification of failure causes. However, as manufacturing systems become increasingly complex, this task has become more difficult, especially for non-experts. Failure cause identification is an iterative process in which possible causes are inferred from currently available information and additional inspections are then performed based on the inference results. This study proposes a method to support non-experts in this process. The method is based on a knowledge base that integrates a Function-Behavior-Structure (FBS) model of the target system with causal chains of failures derived from maintenance records. Given user-entered inspection findings, the method infers candidate failure causes using embedding-based similarity retrieval and provides guidance on the next elements to inspect based on the inference results and the FBS model. The method was implemented as a support tool and evaluated through a case study using a learning factory for automatic LEGO car assembly. The results showed that participants using the tool reached path lengths equal to or longer than those of participants without the tool in all trials, suggesting the effectiveness of the proposed method for supporting iterative failure diagnosis by non-experts.

Keywords: Discrete Event Dynamic Automation Systems, Optimization and Optimal Control, Building Automation
Abstract: Elevator group control (EGC) requires the real-time allocation of calls and scheduling of multiple cars under heterogeneous and time-varying passenger demand while balancing multiple service objectives such as waiting time and congestion. We propose an automated data-driven EGC framework that directly infers demand structure from historical call data by clustering time periods using Wasserstein distances between call matrices. For each identified cluster, a simulation dataset is generated using a high-speed EGC simulator, a surrogate response surface is constructed for multiple performance responses, and cluster-specific optimal control parameters are obtained through desirability-based scalarization and simulation-driven search. This unified pipeline systematically addresses demand heterogeneity across time periods and enables robust, multi-response control parameter optimization without relying on predefined traffic modes or manually segmented time periods. The framework has been developed in collaboration with an elevator manufacturer and is currently being prepared for integration into a real building control system, enabling forthcoming field-level validation in operational environments.

Keywords: Failure Detection and Recovery, Machine learning, Factory Automation
Abstract: The standard autoencoder (AE) baseline for anomalous sound detection aggregates frame-level reconstruction errors by their mean, which can dilute localized anomalies in non-stationary machine sounds. We systematically compare four aggregation functions — mean, max, percentile-95, and variance — across all four machine types in the MIMII dataset at 6~dB SNR. Variance aggregation achieves the highest AUC for both non-stationary types (valve: 0.854, slider: 0.923), while mean aggregation remains best for both stationary types (fan: 0.918, pump: 0.842). The ranking of aggregation functions approximately reverses between the two groups, with variance and mean swapping positions at both ends. We further show that detection performance is more sensitive to the choice of evaluation-time aggregation than to the training loss function, suggesting that a standard mean-trained AE combined with signal-dependent aggregation is a practical and effective strategy.

Keywords: Intelligent Transportation Systems, Tendon/Wire Mechanism, Motion and Path Planning
Abstract: Cooperative cable-driven transportation with mobile robots connects mobile robots to the payload with cables, rather than attaching mobile robots to the payload. A cooperative cable-driven transportation system can move over low obstacles and reduce the mutual influence of mobile robots through flexible connections. In a large-scale environment, a cooperative cable-driven transportation system could be regarded as a point and sophisticated path planning approaches are available. However, in a small-scale environment with space constraints and obstacles, a cooperative cable-driven transportation system cannot be regarded as a point and the simultaneous path and motion planning of the system is a changeling problem. To this end, this paper proposes a novel Reinforcement Learn- ing (RL)-based planning approach for cooperative cable-driven transportation. The paper also develops sampling-based and optimization-based planning approaches based on the existing planning approaches for Cable-Driven Parallel Robots (CDPRs). The paper applies these planning approaches to cooperative cable-driven transportation scenarios in simulation and in the real world. The time consumption of these planning approaches and the quality of planning results are compared and analyzed. Based on the comparison of these planning approaches, the paper provides a guideline for the selection and development of planning approaches for cooperative cable-driven transportation with mobile robots.

Keywords: Learning and Adaptive Systems, Deep Learning in Robotics and Automation, Reinforcement
Abstract: Robotic cloth untangling requires progressively disentangling fabric by adapting pulling actions to changing contact and tension conditions. Because large-scale real-world training is impractical due to cloth damage and hardware wear, sim-to-real policy transfer is a promising solution for this task. However, cloth manipulation is highly sensitive to interaction dynamics, and policies that depend on precise force magnitudes often fail after transfer because similar force responses cannot be reproduced due to the reality gap. We observe that untangling is largely characterized by qualitative tension transitions rather than exact force values. This indicates that directly minimizing the sim-to-real gap in raw force measurements does not necessarily align with the task structure. We therefore hypothesize that emphasizing coarse force-change patterns while suppressing fine environment-dependent variations can improve robustness of sim-to-real transfer. Based on this insight, we propose Adaptive Force-Difference Quantization (ADQ), which reduces observation resolution by representing force inputs as discretized temporal differences and learning state-dependent quantization thresholds adaptively. This representation mitigates overfitting to environment-specific force characteristics and facilitates direct sim-to-real transfer. Experiments in both simulation and real-world cloth untangling demonstrate that ADQ achieves higher success rates and exhibits greater robustness in sim-to-real transfer than policies using raw force inputs.

Keywords: Learning and Adaptive Systems, Industrial and Service Robotics, Calibration and Identification
Abstract: Industrial robots are widely deployed across diverse and complex operational environments due to their flexibility and large workspace characteristics. However, when the industrial robots move significantly within its workspace, their trajectory accuracy degrades considerably. Traditional calibration and error compensation methods, whether model-based or data-based, are typically workspace-specific and collect data only within a small region of the overall workspace. When the range of motion and displacement of an industrial robot expands, a large amount of recalibration data must be acquired for the new motion range, and the collection of such data is costly and time-consuming. To address this challenge, this paper proposes a cross-workspace error compensation framework for industrial robots, which combines a Temporal Convolutional Network (TCN) with transfer learning. Physical experiments are conducted on a standard six-degree-of-freedom industrial robot to validate the performance of the proposed framework across these four distinct workspace and to demonstrate the effectiveness of the method. The position error results in the target workspace after fine-tuning are analyzed to obtain the optimal TCN-TL error compensation model, and further analysis confirms the efficiency and robustness of the proposed method.

Keywords: Manipulation Planning, Compliance and Impedance Control, Reinforcement
Abstract: Articulated-object manipulation is difficult when a robot must follow blind end-effector commands without reliable vision or force sensing, since direct tracking can push into blocked directions and cause jamming or penetration. We present a proprioception-only framework for constraint-aware manipulation that infers local motion constraints online and converts blind commands into contact-compatible motions. The method represents the local constraint normal as a multi-modal belief on the unit sphere using a mixture of von Mises--Fisher distributions predicted from short proprioceptive histories. Belief entropy provides a confidence signal that drives an adaptive soft projection, suppressing only the constraint-violating component of the command while preserving motion needed to maintain contact. A downstream policy trained with Reinforcement Learning (RL) tracks the projected command, and the estimator and policy are jointly optimized during training. Simulation studies in Isaac Lab on drawer and cabinet-door opening under RL-generated and SpaceMouse commands show improved robustness over blind tracking baselines.

Keywords: Manipulation Planning, Motion and Path Planning, Machine learning
Abstract: Motion planning for a pick-and-place task can be formulated as start-to-goal-region problems, in which the multiple grasp and place poses that are valid to complete the task can be represented as workspace goal regions. However, when objects must be manipulated in constrained environments such as shelves, the planner may struggle to rapidly find a path to any of the available goal regions while navigating the complex robot configuration space. Motivated by this challenge, we propose a learning-informed sampling-based tree planner, which increases planning efficiency and the likelihood of finding feasible solutions to such start-to-goal-region problems. Our proposed planner leverages a transformer model to generate configuration samples, which are conditioned on the start configuration, goal configurations, and visual observations of the environment. Such generated configurations are used to help guide the expansion of a sampling-based tree planner, while also maintaining stochastic behavior. We evaluate our planner in both simulated and real-world scenarios of increasing complexity, including shelf environments with a 7-DoF Franka Emika robot. Across all settings, the proposed approach improves planning efficiency while maintaining or increasing success rates. In the most challenging scenarios, our proposed approach increases success rates by over 20% and reduces average planning time by more than a factor of two compared to a goal-region–aware sampling-based tree planner baseline. We also demonstratr effective generalization from simulation to real-world execution.

Keywords: Manipulation Planning, Reinforcement, Actuation and Joint Mechanisms
Abstract: Robotic manipulation faces fundamental challenges in achieving robust grasping across geometrically diverse objects. This study presents VacGrasp, a reinforcement learning framework combining Proximal Policy Optimization (PPO) with a new 4 x 4 modular vacuum gripper system featuring independent z-axis retraction. Our PyBullet simulation environment models suction dynamics through Cartesian force application when gripper-object proximity falls below a threshold value. The policy learns optimal grasping strategies from RGB-D camera observations and a multi-component reward function encoding grasp success and gripper-object Euclidean distance. We evaluated the performance of the learned policy both on seen and unseen objects. Experimental evaluation demonstrates strong generalization capabilities, with the trained policy achieving 97.5% success on flat geometric primitives (boxes/cylinders), 84% on seen complex objects, while maintaining 83% success on unseen objects with moderate curvature and 60% on high-curvature unseen objects. These results reveal a clear correlation between surface curvature and grasping performance, showing that learned policies can effectively transfer vacuum grasping strategies to novel scenarios. The system's ability to handle both seen and unseen objects while adapting to varying surface geometries represents a significant advance toward robust robotic grasping in real-world applications.

Keywords: Factory Automation, Inventory Management, Remanufacturing
Abstract: Maintenance organizations in manufacturing try to avoid downtime and unnecessary purchasing by reusing existing assets, but the main obstacle is not a lack of parts but a lack of actionable visibility across sites and partners. Inventories are distributed, described with inconsistent naming conventions, and contain duplicates and partially specified references, so the right part often exists somewhere but remains effectively undiscoverable. The paper proposes PhRAG, a hybrid Retrieval-Augmented Generation for pooling this fragmented landscape into a Virtual Stock Pool (VSPool) that can be structured and searched as a single resource. Heterogeneous spare part descriptions are structured via Named Entity Recognition (NER) into a shared virtual pool dataset and indexed to support robust retrieval even when users express needs in natural language rather than exact technical specifications. The proposed modular pipeline leverages the multitasking nature of generative language models to cover two dimensions that make industrial parts pooling challenging: (i) unstructured technical specifications from diverse data sources (e.g. new partners, catalogs, marketplace listings) are handled through an offline extraction and (ii) request variability at runtime (references, partial references, specifications, price/condition constraints) is handled through a hybrid RAG-based search engine capable of retrieving relevant components and justifying results. The framework demonstrates the potential of generative approaches compared with traditional NER approaches in the presence of data scarcity for technical specifications extraction and overcomes the opacity of standard information retrieval systems by generating justifications for retrieved components.

Keywords: Intelligent Transportation Systems, Computer Vision for Transportation
Abstract: ，现有的低光增强 6041;法可以提升全ह 6;效果 它们经常难以增 4378;背景音 噪声、高频车道 2454;节模糊，以及ल 8;致 交通场景中的局 7096;过曝（眩光）Ӎ 0;要处理 这些问题，包括GDI M-SAM-SCI，一种改进的&# 33258;校准 照明（SCI）网络， 专门为 夜间车道线图像 2686;强。首先，我ߤ 4;介绍 一个门控细节注 0837;模块（GDIM），采&# 29992; 可学习的软阈值 6426;制以自适应获এ 1; 功能，有效抑制 0302;幅度噪声 在有选择性地增 4378;结构的同时，ຓ 9;用渐变 车道线的详细信 4687;。第二，空间ࠥ 1;注 机构（SAM）集成在 照明中 用于引导道路语 0041;网络的估计分ਞ 3; 区域，防止对非 0844;路区域的

Keywords: Intelligent Transportation Systems, Computer Vision for Transportation
Abstract: Robust 3D Single Vehicle Tracking (SVT) is a pivotal technology for autonomous vehicle following in Unmanned Ground Vehicles (UGVs). However, tracking stability in off-road environments is often compromised by severe terrain undulations and dust-induced occlusions. To address these challenges, this paper proposes a novel cross-modal hierarchical fusion framework—CM-SVT. The framework first processes data from LiDAR, millimeter-wave radar (MMW radar), and cameras to obtain independent object detection or tracking results from each sensor. Subsequently, a Hierarchical Multi-modal Data Association (HMDA) strategy is constructed to facilitate data fusion. By fully leveraging the distinct detection characteristics of heterogeneous sensors, this strategy effectively enhances target tracking continuity in complex off-road scenarios. Furthermore, the framework incorporates IMU measurements into an Extended Kalman Filter (EKF) to achieve precise target state estimation within a global coordinate system. Experimental results across various typical off-road scenarios demonstrate that the proposed method exhibits strong immunity to light-to-moderate dust or smoke, ensuring stable and persistent vehicle-following performance with good robustness. Moreover, the framework maintains high real-time efficiency, achieving a processing frame rate of more than 40 Hz.

Keywords: Optimization and Optimal Control, Software, Middleware and Programming Environments, Logistics
Abstract: Reliable autonomous mapping, environmental sampling, last-mile logistics, and infrastructure deployment depend on the optimal surface area-balanced distribution of Spatial Reference Sites (SRS). Conventional 2D Euclidean methods often fail in high-relief environments by neglecting topographic variations and physical obstructions. This leads to significant planimetric distortion, spatial clustering, and the placement of targets in inaccessible or shadowed regions, compromising both data integrity and operational safety. This paper introduces TOPOS, an automated framework for TOPography-aware Optimal Sampling on topographic manifolds. We treat the terrain as a discrete 2-dimensional manifold embedded in 3D Euclidean space and replace standard flat-map distances with non-Euclidean geodesic distances that follow the actual surface geometry. The point distribution is formulated as an optimization problem using a Constrained Geodesic Voronoi Decomposition, solved via a Riemannian Nesterov Accelerated Gradient (NAG) engine. Our approach restricts target locations to a feasible ”safe zone,” accounting for non-traversable slopes, vegetation, environmental occlusions, etc. Through evaluations on non-convex sinusoidal manifolds, we show that TOPOS mitigates planimetric distortion by utilizing geodesic metrics. This approach results in a ∼74% improvement in optimal surface area-balanced distribution, as measured by the coefficient of variation (CV) of the Voronoi cell areas. The framework is architected as a Geographic Information System (GIS)-ready micro-service to bolster the mentioned applications.

Keywords: Probability and Statistical Methods, Sensor Networks, Learning and Adaptive Systems
Abstract: Modeling of large-scale industrial systems is often challenged by non-stationary dynamics and complex spatial dependencies, which can degrade the performance of traditional static-parameter models. This paper presents a dynamic Gaussian Markov random field-particle filtering (GMRF-PF) framework designed to track time-varying spatial interactions and observation noise as latent states. By formulating the problem within a sequential Bayesian inference process, the model is capable of adapting to evolving spatio-temporal dynamics. Evaluations involving both simulations and real-world grain storage data suggest that the framework offers improved tracking accuracy and robustness over static and sliding-window baselines. Furthermore, the results indicate that the proposed approach facilitates the decoupling of complex spatial structures, providing a reliable foundation for online anomaly detection in industrial infrastructures.

Keywords: Reinforcement, AI-Based Methods, Task Planning
Abstract: In Decentralized Training and Decentralized Execution (DTDE) for cooperative Multi-Agent Reinforcement Learning (MARL), action-advising-based knowledge sharing promotes interpretable and scalable cooperation among agents. However, current action advising approaches often adhere too much to the teacher's guidance without evaluating teacher-student compatibility, which causes excessive advising, suboptimal stability, and degraded performance. To overcome these challenges, this paper presents a Consensus-based Communication and Knowledge Sharing (CCKS) framework, which allows agents to adopt recommendations based on consensus-derived constraints and to follow the teacher's instructions more smartly. This mechanism enables agents to balance exploration and learning from experienced teachers, improving overall performance. The key is the consensus model construction, for which we propose to employ contrastive learning to construct consensus models based on local observations in the agents' training phase. In action selection, agents score and choose actions based on consensus and shared knowledge. Designed as a plug-and-play solution, CCKS integrates seamlessly with existing DTDE algorithms. Experiments conducted in the Google Research Football environment and the complex StarCraft II Multi-Agent Challenge demonstrate that the integration with CCKS significantly improves cooperation efficiency, learning speed, and overall performance compared with current DTDE baselines. The code is available at https://github.com/yuanxpy/CCKS.

Keywords: Collaborative Robots in Manufacturing, Human Factors and Human-in-the-Loop, Telerobotics and Teleoperation
Abstract: Dexterous hands for intelligent manufacturing are being extensively studied with the development of humanoid robot technology. Nevertheless, fine-grained human-robot collaborative dexterous hand tool operations remain challenging because they require precise control, while teleoperation imposes a high burden on the operator during precise approach and grasping. Even worse, learning-based methods may fail when real-world conditions deviate from the training distribution. To address these challenges, this study introduces cone-constrained gated human-in-the-loop shared control for dexterous hand tool manipulation based on a 3D diffusion policy (DP3), ConeGate-DP3. First, the study designed a bilateral teleoperation exoskeleton with wrist force feedback for shared control in the system. This teleoperation exoskeleton not only enables human experts to complete global approach and anomaly handling but also conveys the robot’s intention to the expert through a dedicated force feedback device. Second, the DP3 policy in the system takes over during local phases once both the operator and the system confirm takeover feasibility to accomplish critical fine-grained tool operations, thereby balancing robustness and operational consistency. Third, this study proposes a geometry-based method to assess takeover feasibility from the relative configuration between the end-effector and the target tool. Validation in real-world human-robot collaborative dexterous hand tool handling demonstrates the effectiveness of the proposed system, which combines the advantages of human global decision-making, tactile guidance, and the fine-grained local control of diffusion strategies, making a significant contribution to the field of human-in-the-loop shared control and dexterous tool manipulation.

Keywords: Collaborative Robots in Manufacturing, Human-Centered Automation, Industrial and Service Robotics
Abstract: As a key paradigm of Human-Robot Interaction (HRI) in the Industry 4.0 era, Human-Robot Collaboration (HRC) has become increasingly important in industrial operations such as warehousing and logistics. However, the “silent behavior” of robots that lack functional social cues often impacts the individuals’ trust towards them, affecting the quality of collaboration and causing insecurity of personnel in the long run. Existing research mostly emphasizes the optimization of robot performance but pays little attention to making robots communicate with users. Under the human-centric vision of Industry 5.0, this study proposes a multimodal (audio-visual) functional social cue system grounded in the Computers Are Social Actors (CASA) theory and Proxemics theory to convert robots’ changes in distance and direction into direct functional social cues, thereby improving the transparency and predictability of robots’ actions. Through a within-subjects repeated measures experiment, the results demonstrate that participants’ trust in robots under the condition with the multimodal functional social cue system is significantly higher than their trust under baseline condition (p = 0.007). Findings of the study indicate that the multimodal functional social cue system can effectively enhance trust perception in HRC, which provides a practical, implementable interaction design for industrial human-robot collaboration.

Keywords: Compliant Joints and Mechanisms, Industrial and Service Robotics, Robotics and Automation in Life Sciences
Abstract: Conventional rigid grippers struggle to reliably grasp thin, low-stiffness objects such as cards, coins, and metal sheets due to their inability to adapt to object geometry. Active soft grippers offer adaptability but introduce complexity, cost, and control latency. Existing passive adaptive mechanisms attempt to address these challenges, yet they suffer from structural complexity, sensitivity to assembly tolerances, and susceptibility to lateral deformation under load. This paper presents the X-Claw gripper, a passive adaptive robotic finger based on the Slotted Eight-Linkage (SEL) mechanism. The SEL mechanism integrates two nested parallelograms with a slotted rail-slider constraint, inherently constraining the fingertip to a precise vertical trajectory without requiring additional parallelogram components. This novel configuration eliminates the structural complexity and assembly sensitivity found in conventional passive linkages, while the slotted constraint ensures rigid motion transmission free from lateral deformation. The gripper achieves passive mode switching between pinching and scooping through environmental interaction alone, enabled by a beveled fingertip geometry and an extension spring for energy storage and posture recovery. Experimental validation demonstrates the X-Claw’s effectiveness in pinch grasping, thin-object scooping, and adaptive grasping in cluttered environments, offering a low-cost, sensorless solution for reliable thin-object manipulation.

Keywords: Compliant Joints and Mechanisms, Industrial and Service Robotics, Robotics and Automation in Life Sciences
Abstract: This paper presents an underactuated gripper (PLP Gripper) capable of achieving stable pinching operations under various environmental constraints. The fingertip of the gripper can adapt to unstructured environments and realize a gripping-lifting effect on lightweight objects. The mechanical gripper consists of two finger phalanxes, integrating dual-parallelogram mechanisms and pulley-link mechanisms. The dual-parallelogram linkages ensure that the fingertip maintains a fixed vertical posture relative to the base, while the pulley-linkage mechanisms provide the device with adaptability to object surfaces and a lifting effect. By incorporating elastic elements and limit blocks, the system preserves the initial posture of the finger, preventing deviations in the initial grasping position caused by finger deformation in non-operational states and thereby ensuring the stability of the grasping process. Experimental results verify the gripper's exceptional adaptive and gripping-lifting performance, demonstrating its broad application potential in complex dynamic environments.

Keywords: Compliant Joints and Mechanisms, Mechanism Design in Meso, Micro and Nano Scale, Manipulation Planning
Abstract: Traditional robotic hands with parallel or coupled linkages often suffer from curved fingertip trajectories and unstable posture during grasping, making it difficult to manipulate thin or flat objects under environmental constraints. To overcome these limitations, this paper introduces the DPS (Dual-Parallelogram and Sliding-Guide) hand, a novel robotic gripper that integrates a dual-parallelogram linkage system with a slot-guided sliding mechanism. The combination of the dual-parallelogram structure and the slot-guided mechanism enables the fingertip to achieve a stable and highly linear trajectory, supporting precise and consistent grasping motions. By coordinating these mechanisms with passive elastic elements, the DPS Hand achieves linear-parallel pinching, self-adaptive grasping, and posture-preserving environmental interaction without the need for complex control strategies. A complete kinematic model is developed to characterize the fingertip trajectory and analyze the motion transmission conditions in both linear and adaptive phases. A 3D-printed prototype is fabricated for experimental evaluation, and the results demonstrate that the DPS Hand can reliably grasp thin planar objects, irregular shapes, and multi-sized items with high stability and compliance. The proposed DPS architecture offers a compact and mechanically efficient solution for grasping tasks in both structured and unstructured environments.

Keywords: Cyber-physical Production Systems and Industry 4.0, Simulation and Animation
Abstract: Radio frequency (RF) signals are emerging as a critical sensing modality for resilient perception in industrial automation and autonomous mobile robots (AMRs), offering non-line-of-sight (NLoS) capabilities via multipath propagation. However, the advancement of RF-aware automation is hindered by the lack of a unified Digital Twin capable of bridging high-fidelity robot dynamics with complex electromagnetic environments. Existing workflows rely on grid-based decoupled simulations that fail to capture the cyber-physical synchronization required for active, closed-loop navigation.

To address this, we propose an open-source RF-Robot Digital Twin that tightly couples rigid-body physics (Isaac Sim) with site-specific ray tracing (Sionna RT). Driven actively by standard ROS navigation stacks, the framework utilizes an asynchronous latest-available scheduling architecture to mitigate the timescale mismatch between fast physics updates and computationally intensive RF modeling. This design maintains a Real-Time Factor (RTF) >1, guaranteeing operational stability for Hardware-in-the-Loop (HIL) testing even under Massive MIMO workloads (up to 4096 antenna pairs). We validate the twin’s sim-to-real fidelity against a physical hardware testbed, achieving a 0.35^{circ} Mean Absolute Error in continuous dynamic Angle-of-Arrival tracking and a 0.847 spectral cosine similarity. By synthesizing synchronized streams of kinematics, vision/LiDAR, and RF channel observations, this platform provides a verified cyber-physical testbed for the design and deployment-readiness assessment of RF-aware automation systems.

Keywords: Motion and Path Planning, Human-Centered Automation, Computer Vision in Automation
Abstract: Evaluating robots in shared human environments requires navigation tasks that capture realistic interaction structure. We present a pipeline for constructing shared-space navigation scenarios from continuous real-world pedestrian trajectory data. Candidate temporal windows are mined using an interaction salience score combining proximity events, local density, and heading diversity. For each retained window, deterministic start–goal pairs are constructed relative to an inferred interaction locus and dominant motion axis, producing traversal tasks that pass through regions of human activity while respecting static geometry. A geometry-constrained A* traversal is used as a feasibility check and reference path. We further introduce auditing procedures to characterize selection bias, endpoint degeneracy, and geometric connectivity. Experiments on a month-long retail trajectory corpus demonstrate that the resulting scenarios are heterogeneous, interaction-rich, and reproducible. The deployed pipeline provides a reproducible substrate for evaluating shared-space robot planners and prediction-informed navigation policies using real-world human motion structure.

Keywords: Motion and Path Planning, Model Learning for Control, Collision Avoidance
Abstract: Conventional snake-like manipulators are plagued by limited mobility, inadequate flexibility, low load-bearing capacity, and inefficient motion planning when operating in constrained 3D environments. To address these challenges, this paper proposes a rigid-flexible coupled mobile snake-like manipulator along with an enhanced diffusion policy-based follow-the-leader motion planning method. The proposed manipulator integrates an all-terrain mobile chassis, a three-joint rigid manipulator, and a seven-joint flexible manipulator actuated by a stepped double-screw cable system; it adopts rigid inverse kinematics and flexible hierarchical kinematics to achieve high-precision trajectory tracking. Derived from the Denoising Diffusion Probabilistic Model (DDPM) framework, the planning method fuses PointNet-based scene encoding, cross-attention feature fusion, and trajectory post-processing techniques (including endpoint correction, trajectory pruning, and smoothing), which collectively enhance its adaptability to environmental constraints and engineering practicability. Experimental results confirm the integrated system exhibits outstanding performance in load-bearing capacity, motion accuracy, and trajectory tracking. Specifically, the proposed method attains a 79.6% planning success rate, outperforming state-of-the-art baselines by a margin of 10%-20%, while delivering optimized trajectory length, obstacle clearance, and smoothness. This work offers an efficient and robust solution for motion planning of confined-space robots, and advances the application of generative AI in bionic motion planning.

Keywords: Motion Control, Optimization and Optimal Control, Motion and Path Planning
Abstract: Achieving safe quadrupedal locomotion in real-world environments has attracted much attention in recent years. When walking over uneven terrain, achieving reliable estimation and realising safety-critical control based on the obtained information is still an open question. To address this challenge, especially for low-cost robots equipped solely with proprioceptive sensors (e.g., IMUs, joint encoders, and contact force sensors), this work first presents an estimation framework that generates a 2.5-D terrain map and extracts support plane parameters, which are then integrated into contact and state estimation. Then, we integrate this estimation framework into a safety-critical control pipeline by formulating control barrier functions that provide rigorous safety guarantees. Experiments demonstrate that the proposed terrain estimation method provides smooth terrain representations. Moreover, the coupled estimation framework of terrain, state, and contact reduces the mean absolute error of base position estimation by 64.8%, decreases the estimation variance by 47.2%, and improves the robustness of contact estimation compared to a decoupled framework. The terrain-informed CBFs integrate historical terrain information and current proprioceptive measurements to ensure global safety by keeping the robot out of hazardous areas and local safety by preventing body-terrain collision, relying solely on proprioceptive sensing.

Keywords: Planning, Scheduling and Coordination, Robot Networks, Optimization and Optimal Control
Abstract: Multi-robot exploration in cluttered, unknown environments depends on inter-robot communication and mutual observation, yet occlusions and nonholonomic motion can easily break line-of-sight (LoS) and fragment the team. This paper investigates distributed exploration with multiple nonholonomic robots under LoS connectivity constraints. We present a closed-loop exploration framework that couples frontier-driven goal selection with LoS-constrained planning. We maintain team connectivity by building a weighted minimum spanning tree online based on visibility and communication distance. This produces robot neighborhood relations that are necessary for enforcing LoS. Each robot then solves a distributed, optimization-based planning problem using only onboard sensing and information from its neighbors. This produces efficient and dynamically feasible trajectories while satisfying collision avoidance and LoS constraints to the robot’s current neighbors. We validate our approach in simulation and on hardware, demonstrating improved exploration continuity and reduced team fragmentation in challenging environments.

Keywords: Reactive and Sensor-Based Planning, Cloud Computing For Automation, Compliant Assembly
Abstract: Recent advances in low-cost manipulation platforms have brought mass adoption of consumer robots closer to reality, yet many daily tasks (e.g. plugging in appliances and charging cables) require precise, contact-rich manipulation that demands accuracy approaching the limits of inexpensive platforms. We posit that the fundamental limitation is nonetheless not hardware alone, but rather a mismatch between existing approaches (that require accurate sensing or low-latency perception-to-action pipelines) and the reality of these platforms, and that belief-space planning offers a path forward. We present a belief-aware policy planning framework built on two technical contributions: (1) a dual-setpoint contact-switching policy optimized in belief space, in which belief-space planning simultaneously optimizes two velocity trajectories corresponding to free-space and contact modes while accounting for the unreliability of motor-current-based contact detection; and (2) a belief-constraining kinematic extension that can be actively positioned to contact the environment, reducing uncertainty. We integrate these with a stereo depth estimation pipeline using a pair of basic RGB cameras, demonstrating that our belief-aware planner can accommodate inaccurate pose estimates from inexpensive hardware. An important practical benefit of our approach is a complete decoupling of planning from execution while remaining partially closed-loop: computation runs on remote infrastructure without latency constraints, while the local system only executes simple PD setpoint commands derived from velocity specifications and mode switching. We validate our approach on a modified SO-101 platform, demonstrating on average 88.3% success in real-world connector insertion (discounting perception failures).

Keywords: Reactive and Sensor-Based Planning, Computer Vision in Automation, Autonomous Agents
Abstract: We present an autonomous system for 3D reconstruction of bounded objects using a UAV equipped with an RGB-D camera. The system combines incremental 3D Gaussian Splatting (3DGS) with a novel depth-variance acquisition function for next-best-view (NBV) planning. Unlike alpha-based coverage metrics that are fooled by hallucinated renders, our scorer measures per-pixel geometric uncertainty via alpha-blended depth second moments in a single render pass. We demonstrate the system on a PX4 drone in Gazebo simulation, achieving 73% surface coverage compared to 40% for fixed orbit patterns, with comparable reconstruction quality. The depth-variance signal maintains viewpoint differentiation throughout the mission while conventional alpha coverage saturates, enabling informed viewpoint selection even with mature models.

Keywords: Planning, Scheduling and Coordination, Semiconductor Manufacturing, Task Planning
Abstract: The Parallel-Batch Scheduling Problem on Unrelated Parallel Machines (PBSP-UPM) with serial setup times is a complex combinatorial optimization problem frequently encountered in modern manufacturing systems, such as additive manufacturing. While metaheuristics are widely used in solving similar scheduling problems, the influence of serial setup time on PBSP-UPM is rarely considered. Consequently, Mixed-Integer Linear Programming (MILP) formulations remain open. The MILP models formulated according to similar scheduling problems are computationally prohibitive for practical instances. The primary bottleneck lies in the severe symmetry and combinatorial explosion caused by traditional global index-based constraints, which lead to poor linear programming relaxations and massive redundant branches. To overcome this limitation, this paper proposes a novel and highly efficient MILP model to minimize the makespan. Our core contribution is a-priori local job index bijection mechanism that reconstructs the constraint space. Instead of relying on rigid global indices, the proposed model aligns the symmetry-breaking logic directly with the machine-specific processing times. This exact pre-processing procedure drastically reduces the constraint spatial complexity of the model, eliminating nearly 50% of the constraints compared to the baseline model, while keeping the number of variables identical. Extensive computational experiments across datasets with varying batch composition complexities demonstrate that the proposed model significantly outperforms the baseline model in computational speed, exact solvability, and robustness against severe combinatorial explosion.

Keywords: Swarms, AI-Based Methods, Machine learning
Abstract: In this paper, an Enhanced Traffic Jam Optimizer (ETJO) is proposed to solve the problems that the original Traffic Jam Optimizer (TJO) is prone to premature convergence and late search stagnation when solving complex constraints and high-dimensional optimization problems. Firstly, the algorithm introduces a dynamic weather perception mechanism to balance global exploration and local exploitation. Secondly, a heavytailed sinusoidal perturbation strategy is integrated to enhance the ability to escape from local optima. Finally, a deviationbased guidance strategy is designed to improve the convergence speed and accuracy. To verify the effectiveness of ETJO, this paper not only evaluates the performance of ETJO on CEC2022 test set, but also applies it to six typical engineering design problems and 12 UCI datasets for feature selection experiments. The experimental results show that ETJO significantly outperforms the comparison algorithm and the original TJO in terms of convergence speed, optimization accuracy and stability. In addition, ETJO can achieve higher classification accuracy with fewer feature subsets in feature selection. In summary, ETJO has strong competitiveness and application value in handling high-dimensional data optimization and complex industrial engineering problems.

Keywords: AI-Based Methods, Machine learning, Big-Data and Data Mining
Abstract: Prompt tuning offers an efficient way to adapt frozen vision-language models to unsupervised domain adaptation (UDA). Existing prompt-based UDA methods usually rely on domain-aware prompts, but they often couple category semantics and domain correction throughout the prompt hierarchy. This coupling can introduce redundant class encoding in deep prompts and limit the specialization of prompts for domain shift. We propose HDPT, a hierarchically decoupled prompt tuning framework for CLIP-based UDA. HDPT uses shallow prompts to anchor category semantics and domain identity, while deeper prompts discard explicit class prompts and focus on progressive domain correction. To couple the two modalities, shallow textual prompts are aggregated and projected to initialize the visual prompts, and layer-wise text-to-vision mappings are used in deeper layers. We further align source-conditioned and target-conditioned predictions with a Pearson correlation loss, which preserves cross-domain class-relation consistency without enforcing identical probability calibration. Experiments on Office-Home, VisDA-2017, mini-DomainNet, and Office-31 show that HDPT achieves competitive performance against recent prompt-based UDA methods, with clear gains on style-diverse transfer tasks. Code is available at https://github.com/zekaizhou/HDPT.git.

Keywords: Optimization and Optimal Control, Model Learning for Control, Process Control
Abstract: Cold rolling is a key process in steel manufacturing. In this process, optimal batching and scheduling plans are essential for minimizing production cost. While single-campaign scheduling has been well-studied, multi-campaign coordination remains largely underexplored. In this paper, we formulate the cold rolling multi-campaign batching and scheduling problem as a prize-collecting vehicle routing problemand establish an integer programming model to minimize the total production cost, comprising campaign opening cost, sequence-dependent transition penalties, and order rejection penalties. To solve the model, a hybrid ant colony optimization algorithm is proposed, which constructs blocks from low-cost node paths and employs a multi-armed bandit strategy to adaptively select local search operators. Computational experiments based on real-world industrial instances show that the proposed algorithm significantly outperforms conventional heuristics in terms of solution quality and stability for large-scale problems.

Keywords: AI-Based Methods, Learning and Adaptive Systems, Reinforcement
Abstract: This paper addresses the challenge of maintaining deterministic safety and tracking performance in non-stationary industrial systems using a Digital Twin (DT) framework. We propose a forward-invariant tube-based architecture for intelligent optimization that bridges high-level Reinforcement Learning (RL) with rigorous industrial stability requirements. To mitigate the Sim-to-Real mismatch in complex manufacturing environments, we utilize a Projected Recursive Least Squares (RLS) estimator to establish a certified error bound. Our 'Safe-by-Design' shield integrates Lyapunov-IQC theory with an Actor-Critic framework via a convex Quadratic Program (QP) filter, ensuring hard invariance by confining control signals within a certified ellipsoidal tube. This eliminates chattering, which is critical for preserving mechanical integrity in precision equipment. Theoretical analysis confirms the system is Uniformly Ultimately Bounded (UUB). Validated on robotic benchmarks, the method achieves zero safety violations and an 84.7% improvement in tracking precision compared to standard industrial PID baselines, demonstrating high potential for optimizing complex industrial equipment.

Keywords: Sustainability and Green Automation
Abstract: The adaptability and intelligence of flexible manufacturing systems are key characteristics for responding to rapid changes in personalized market demand. Large Language Models (LLMs) have demonstrated outstanding performance in accomplishing complex planning tasks and can provide a new paradigm for the adaptive optimization of flexible manufacturing systems. Thus, a large language model-driven intelligent decision-making method is proposed for flexible process planning and cutting parameter optimization. Firstly, we design three core agents based on the LangGraph and LangChain frameworks to simulate flexible process planning and cutting parameter optimization. Then, a data synthesis framework for simulating the interaction of manufacturing systems is developed, and the synthesized data is used to train the LLMs to accurately handle manufacturing process optimization. Finally, experiments are conducted to illustrate the workflow of flexible process planning and cutting parameter optimization and to verify the effectiveness of the proposed optimization method. The case study shows that the proposed method exhibits strong adaptability in dealing with various optimization tasks. In addition, the proposed method can be integrated with different types of tools to adapt to diverse machining tasks.

Keywords: Health Care Management, Scheduling in Healthcare, Modelling, Simulation and Optimization in Healthcare
Abstract: Many healthcare services are delivered through consecutive multi-visit appointment courses rather than isolated single visits. In such settings, each patient is assigned a temporally linked sequence of appointments, thus a failure to initiate the treatment may invalidate the entire pre-reserved block of future capacity rather than a single slot. As a result, the scheduling process becomes inherently block-based, with limited rescheduling flexibility and strong intertemporal coupling in capacity decisions. To address this challenge, we develop an online stochastic optimization model while still preserving the block allocation structure of consecutive treatment courses. Numerical experiments compare the proposed model with two benchmarks: a conventional hospital scheduling rule and an offline global optimization policy. The latter serves as a full-information upper bound on achievable performance. The results indicate that the proposed model substantially reduces waiting-to-start time and cost relative to the conventional rule, with only moderate increases in within-day waiting and overtime, although it remains below the offline global benchmark because of the short-sightedness of online decisions. Additional experiments further characterize the model under different system conditions and weight settings, and illustrate the trade-offs among patient access, congestion, and operational burden. Such a study provides a practical online stochastic optimization framework for course-based appointment scheduling under treatment-initiation dropout as well as the structural properties and trade-offs.

Keywords: Health Care Management, Scheduling in Healthcare, Modelling, Simulation and Optimization in Healthcare
Abstract: Outpatient appointment scheduling critically affects patient waiting time and resource utilization in healthcare, but stochastic patient behavior complicates the prediction of three core metrics: the average waiting time of all patients (AWT), the average waiting time of punctual patients (P-AWT), and physician idle time (IDT). To address the challenge, we integrate discrete-event simulation (DES) with a dual-strategy ensemble learning. The DES model is developed based on real-world hospital data in Beijing. An ensemble XGBoost surrogate, with feature engineering to expand the appointment vector, is developed based on the DES model to achieve improved prediction accuracy. A dual-strategy ensemble method enables accurate and rapid prediction of the three core metrics. The results show that the surrogate model with ensemble learning maintains high accuracy in regular scenarios with relative error around 1.3%, and it can also achieve accurate prediction across various abnormal scenarios, with the maximum relative error around 10%. The proposed surrogate model provides fast predictions compared to the simulation, thereby providing efficient decision support for appointment scheduling.

Keywords: Causal Models, AI and Machine Learning in Healthcare
Abstract: The advent of increasingly available patient health data has created opportunities for critical care medicine to embrace modern data-driven decision support. However, existing methods, such as rule-based scoring systems or machine learning-based classification models, often focus on uni-dimensional patient outcomes such as mortality, lacking the granularity necessary to understand a patient’s overall health during their hospital stay. This research focuses on predicting a patient’s clinical trajectory, described by key health indicators readily accessible through electronic health records, by estimating correlations and causal relationships among these health indicators over time. To overcome the challenges associated with multi-dimensional, longitudinal, and heterogeneous patient data, we develop an analysis pipeline that first identifies distinct patient clusters with similar clinical trajectories through temporal pattern analysis. Then, for each cluster, we implement a structural equation model (SEM) to capture both immediate and time-lagged dependencies among variables, which offers a deeper understanding of the dynamic relationships. Directed acyclic graphs (DAGs) are used to visualize causal and correlative relationships, providing intuitive insights into how vital health indicators interact and form a patient’s clinical trajectory. This integrated approach, which combines clustering, SEM, and DAG visualization, enables more refined modeling of patients’ clinical trajectories and has the potential to guide personalized interventions, ultimately improving patient outcomes in critical care settings.

Keywords: AI and Machine Learning in Healthcare, Health Care Management, Modelling, Simulation and Optimization in Healthcare
Abstract: In this study, we present an interpretable multimodal framework that predicts multiple quantiles of surgical duration and constructs case-level predictive distributions from perioperative electronic health records. Evaluation on data sets from two hospitals shows that the proposed approach performs strongly in both point and probabilistic prediction and outperforms benchmark methods in most settings. The framework provides a promising approach for more informative and transparent perioperative decision support.

Keywords: AI and Machine Learning in Healthcare, AI-Based Methods, Clinical and Operational Decision Support
Abstract: Precision oncology seeks to improve cancer care by matching tumor-specific molecular profiles with effective therapeutic strategies. A key component of this process is identifying therapeutically actionable gene interactions that connect molecular alterations to potential treatment vulnerabilities. Synthetic lethality represents an important class of such interactions, where alteration or loss of one gene creates selective vulnerability to disruption of another gene or pathway. This work presents an AI-assisted workflow for extracting synthetic lethal gene pairs from biomedical literature as a practical approach to identifying actionable gene interactions. The proposed framework enables scalable construction of structured evidence that can support clinical decision-making, including the prioritization of therapeutic targets and the design of rational combination treatment strategies.

Keywords: Scheduling in Healthcare, Reinforcement, AI-Based Methods
Abstract: Patients with intracerebral hemorrhage (ICH) under continuous sedation receive frequent neurological assessments and laboratory panels as standard care, yet this standard is increasingly recognized as a source of harm in itself. We formulate adaptive measurement scheduling for sedated ICH patients as an offline constrained Markov decision process (CMDP) on the MIMIC-IV cohort, minimizing cumulative measurement cost while retaining a specified fraction of the information gain (IG) attained by clinician practice. For a given set of measured panels, IG is the leave-one-out change in predicted short-term mortality: each panel's contribution is the absolute difference between the prediction using all measured panels and the prediction when that panel's value is reverted to its decay-imputed prior. These probabilities are supplied by a GRU-D patient digital twin trained on the ICH cohort to predict short-term mortality from irregularly sampled, partially observed time series, and per-panel IG is obtained by leave-one-out inference on this frozen model. The constrained policy is learned by primal–dual optimization of a Lagrangian over two value functions, Q_c for cumulative cost and Q_g for cumulative IG, on top of an IQL-based offline RL framework.

Keywords: Sensor Networks, Discrete Event Dynamic Automation Systems, Planning, Scheduling and Coordination
Abstract: Low Earth Orbit (LEO) constellations provide essential data transmission channels for ground-based edge-cloud computing in 6G Space-Air-Ground Integrated Networks (SAGIN). Routing data for these delay-sensitive tasks requires intersatellite routing to efficiently manage backhaul timeliness and limited onboard energy. However, dynamic topologies and queuing fluctuations cause severe delay uncertainty. Furthermore, state-space explosion in large-scale networks complicates energy optimization under strict delay constraints. To address these challenges, an event-driven learning framework for constrained distributed routing is proposed. First, routing is modeled as a constrained stochastic shortest path (C-SSP) problem with an absorbing terminal state via a constrained Markov Decision Process (CMDP) to minimize energy consumption under delay constraints. Second, an event-driven mechanism quantizes delay and resource slack into discrete events. This formulates a stochastic parameterized policy in a low-dimensional space, significantly reducing decision complexity. Finally, delay constraints are decoupled using Lagrangian relaxation, enabling a two-scale collaborative optimization. This alternates centralized multiplier updates with parallel distributed policy improvements, allowing individual nodes to execute decisions based solely on local observations and Monte Carlo estimations. Simulations demonstrate the method reduces energy consumption while guaranteeing expected delays.

Keywords: Optimization and Optimal Control, Cloud Computing For Automation, Reinforcement
Abstract: With the rapid proliferation of the Industrial Internet of Things (IIoT), the centralization of massive data streams from distributed industrial sites to the cloud has precipitated severe uplink congestion at the network edge. Conventional reactive congestion control mechanisms inevitably suffer from inherent control latency, while static bandwidth allocation strategies fail to accommodate the stochastic bursts characteristic of industrial traffic, resulting in either resource underutilization or severe packet loss. To address these challenges, we propose a time-aware deep reinforcement learning (DRL) framework for proactive macro-bandwidth orchestration in industrial software-defined wide area networks (SD-WANs), leveraging recurrent policy optimization techniques incorporating gated recurrent units (GRUs). By capturing the temporal dependencies of bursty industrial traffic, the agent learns to proactively regulate the transmission rate of SD-WAN edge gateways, mitigating congestion at the capacity-constrained core and safeguarding mission-critical streams against packet loss. Extensive experiments demonstrate that our approach outperforms memoryless DRL and conventional heuristic baselines, reducing sample complexity by more than 50% while constraining the packet loss rate at the cloud hub to 1.20% to achieve the lowest aggregate end-to-end packet loss under heavily saturated conditions.

Keywords: Cloud Computing For Automation, Reinforcement, Sensor Fusion
Abstract: In edge-cloud collaborative computing architectures, when multiple edge servers transmit data to the cloud over a capacity-constrained channel, efficient link resource allocation is crucial to ensure that the cloud receives fresh information. Allocation decisions made in one time slot affect not only current performance but also future system performance, thereby necessitating foresight in the allocation algorithm. To address this challenge, we propose a proximal policy optimization (PPO)-assisted greedy algorithm (PAGA) for link resource allocation. First, based on the age of information (AoI) gain per unit link capacity in the current time slot, a greedy algorithm is employed to evaluate the transmission priorities of newly generated samples and partially transmitted samples in the buffers of sensor nodes served by each edge server. Each sample therefore obtains a greedy score that reflects only the immediate benefit within the current time slot. Subsequently, PPO captures the long-term impact of resource allocation decisions on the overall system AoI and produces a PPO-assisted score for each sample that adaptively refines the greedy evaluation. Finally, the two types of scores are fused and ranked to determine the final transmission priorities, according to which link resources are allocated to each edge server. Simulation results show that the proposed algorithm consistently achieves a lower average weighted AoI than benchmark algorithms.

Keywords: Cloud Computing For Automation, Planning, Scheduling and Coordination, Data fusion
Abstract: In edge computing environments, the service migration decision at the current time slot determines not only the latency and number of migrations in that slot but also those in subsequent time slots. Optimizing migration decisions independently at each time slot cannot capture their impact on future states and thus degrades system performance. To address this challenge, this paper proposes a trajectory prediction and dynamic programming-based service migration algorithm for edge computing, which aims to jointly optimize service latency and the number of migrations. Specifically, a deep learning model is first employed to predict the future trajectory of the user. Based on the predicted trajectory, a finite rolling window is constructed. A dynamic programming algorithm is then applied to obtain the optimal migration decisions within the finite window, and backtracking is used to determine the server selection for the current time slot. The algorithm operates in a rolling optimization manner. At each time slot, the subsequent migration path is replanned according to the latest user location and the updated predicted trajectory, so that the predicted trajectory can be corrected based on the latest location information. Simulation results demonstrate that the proposed algorithm achieves improvements in reducing system service latency and decreasing the number of migrations.

Keywords: AI-Based Methods, Computer Vision in Automation, Machine learning
Abstract: Power transmission lines are vital for stable power grid operations but are threatened by intrusions from foreign objects, such as construction machinery, which can cause significant disruptions. Traditional manual inspections are inefficient and costly, while existing deep learning detection methods often struggle with high false positive and negative rates. To address these challenges, this paper proposes a foreign object intrusion detection method for transmission lines based on large visual models. This method achieves high-precision detection by cascading a fine-tuned Segment Anything Model (SAM) with a Contrastive Language–Image Pre-training (CLIP) model. SAM is enhanced for better segmentation in complex backgrounds, and a prompt library for CLIP enables effective classification of intrusive objects. Additionally, geometric heuristic rules are applied to refine accuracy further. Experimental results on a proprietary transmission line intrusion dataset demonstrate the effectiveness of the approach. The fine-tuned SAM achieves a mean Intersection over Union (mIoU) of 78.51%. With optimized prompts, the CLIP model attains an average classification accuracy of 82.57%. By incorporating geometric heuristic algorithms, the overall average detection accuracy is further improved to 85.23%. This method significantly enhances detection precision for foreign object intrusion in transmission line scenarios—particularly forconstruction machinery—while improving generalization capability in complex backgrounds, providing new insights for intelligent transmission line inspection.

Keywords: Planning, Scheduling and Coordination, Intelligent and Flexible Manufacturing, Reinforcement
Abstract: 旨在解决频繁切换机器的问题， 高能耗和显著的碳排放 在一个灵活的工装车间，采用工程师按订单模式。在 这篇论文，是一份多资源受限灵活的工作 车间排班问题，考虑交货时间 （MRFJSP-DT）被提议。具体来说，模型为 通过积分最大完成时间构建，总计 碳排放与总延误作为多目标 优化目标。为了有效解决这个问题，需要 基于改进的多目标灰狼优化 Q-学习（QL-MOGWO）被开发出来。通过合并 Q-学习以增强局部搜索策略， 算法避免被局限于局部最优解，而 提升解质量以促进最优 安排时间。广泛的实验结果 证明所提出的模型

Keywords: Bioinspired Robot Learning, Collision Avoidance, Sensor-based Control
Abstract: This article presents a two-stage, biologically inspired neuromorphic framework for autonomous robot navigation in cluttered environments. The proposed architecture combines unsupervised spiking neural network (SNN)–based sensory abstraction with reward-modulated spike-timing-dependent plasticity (R-STDP) for decision-making. In the first stage, raw 360° LiDAR measurements are transformed into a compact, interpretable obstacle state through a lateral-inhibition-driven STDP network, yielding a low-dimensional, behaviorally relevant perception of the environment. In the second stage, navigation actions are learned via reward-modulated STDP operating on this abstracted state, supporting long-horizon goal-directed behavior without backpropagation or deep reinforcement learning. Two navigation paradigms are investigated: conventional goal-oriented navigation using relative bearing information, and a probabilistic field–based formulation that enables source-seeking behavior under goal uncertainty. The proposed approach is evaluated extensively in Gazebo and NVIDIA Isaac Sim using a TurtleBot3 platform across static, dynamic, and near-realistic environments. Experimental results demonstrate reliable navigation performance, competitive success and collision rates compared to state-of-the-art SNN and hybrid SNN–RL methods, and substantially lower estimated energy consumption. These findings highlight the effectiveness of modular, biologically plausible

neuromorphic architectures for energy-efficient autonomous navigation in complex environments.

Keywords: Industrial and Service Robotics, Computer Vision in Automation, Task Planning
Abstract: Autonomous stocking in retail environments, particularly supermarkets, presents challenges due to dynamic human interactions, constrained spaces, and diverse product geometries. This paper introduces an efficient modular robotic system for autonomous shelf stocking, integrating commercially available hardware with a scalable algorithmic architecture. A major contribution of this work is the system integration of off-the-shelf hardware and ROS2-based perception, planning, and control into a single deployable platform for retail environments. Our solution leverages Behavior Trees (BTs) for task planning, fine-tuned vision models for object detection, and a two-step Model Predictive Control (MPC) framework for precise shelf navigation using ArUco markers. Laboratory experiments replicating realistic supermarket conditions demonstrate reliable performance, achieving over 98% success in pick-and-place operations across a total of more than 700 stocking events. However, our comparative benchmarks indicate that the performance and cost-effectiveness of current autonomous systems remain inferior to that of human workers, which we use to highlight key improvement areas and quantify the progress still required before widespread commercial deployment can realistically be achieved.

Keywords: Compliance and Impedance Control, Human Factors and Human-in-the-Loop, Model Learning for Control
Abstract: This paper presents a method for user-driven robot Learning from Demonstration (LfD) that reduces user effort while ensuring compliant and precise reproduction. The method eliminates repeated teaching for the same task and enables real-time learning from a single demonstration. Demonstrated motions are reproduced with high precision, while impedance variations are learned in real time to provide both compliance and robustness against perturbations. This mitigates potential safety issues in Human-Robot Interaction (HRI) that arise from conventional time-indexed trajectories lacking compliance. The proposed approach integrates a three-dimensional (3D) Fast Diffeomorphic Matching (FDM) algorithm with a Dynamical System (DS)-based motion generator to achieve real-time single-shot demonstration learning and reproduction. An Extended Kalman Filter (EKF) framework compensates for reproduction errors and recovers from external interactions. Furthermore, an impedance parameterization function is incorporated to learn impedance variations from demonstrations and maintain surface contact for specific applications. The proposed approach is validated through comprehensive experiments on a 7 Degree-of-Freedom (DOF) KUKA LWR IV+ robot.

Keywords: Reinforcement, Deep Learning in Robotics and Automation
Abstract: While many existing grasping models can be highly reliable in picking objects in most cases, challenging scenarios persist in industrial automation where objects are difficult to grasp—such as when positioned in corners, occluded by other items, or tightly clustered. These challenges are prevalent in smart manufacturing and logistics systems, where current robotic systems often require costly human intervention for handling difficult cases. To address this automation gap, we propose a multi-step manipulation approach that combines non-prehensile and prehensile actions through two collaborative policies: a main policy responsible for picking, and a supporting policy that assists by rearranging the scene when the main policy lacks confidence. In particular, we leverage equivariant neural networks to encode SE(2) symmetries in both policies, enabling sample-efficient reinforcement learning from visual observations. Using a symmetry-aware DQN-based approach, we first train the main policy, then train the supporting policy to maximize the main policy's confidence through a value-based reward function. Our method significantly outperforms single-policy approaches, achieving 1.3-2times better performance across five manipulation tasks in simulation. Crucially, we demonstrate efficient on-robot learning directly on hardware (1-2 hours) for three tasks, showcasing the practical viability of our approach for smart industry applications where rapid deployment and adaptation are essential.

Keywords: Motion and Path Planning, Formal Methods in Robotics and Automation, Task Planning
Abstract: We study Dubins-constrained coverage on grids in applications where visiting a cell triggers an irreversible local action, such as painting or drilling. In these settings, re-entering treated cells destroys the initial action and is therefore undesirable, even if doing so would shorten the overall path. However, the motion constraints of a Dubins vehicle force entries into neighbouring cells during turns or when connecting distant regions. We refer to entries into previously treated cells as revisits. Motivated by this, we define the Dubins Coverage with Minimal Revisits (DCMR) problem: find a trajectory that covers every cell while minimising the number of revisits. We derive theoretical bounds on the number of revisits and boundary violations, and complement these results with a configurable solver supporting various decomposition and coverage pattern choices. We then experimentally evaluate how these choices affect the number of revisits and boundary violations.

Keywords: Haptics and Haptic Interfaces, AI-Based Methods, Virtual Reality and Interfaces
Abstract: We present VLH, a novel Visual-Language-Haptic Foundation Model that unifies perception, language, and tactile feedback in aerial robotics and virtual reality. Unlike prior work that treats haptics as a secondary, reactive channel, VLH synthesizes mid-air force and vibration cues as a direct consequence of contextual visual understanding and natural language commands. Our platform comprises an 8-inch quadcopter equipped with dual inverse five-bar linkage arrays for localized haptic actuation, an egocentric VR camera, and an exocentric top-down view. Visual inputs and language instructions are processed by a fine-tuned OpenVLA backbone—adapted via LoRA on a bespoke dataset of 450 multimodal scenarios - to output a 7-dimensional action vector (Vx, Vy, Vz, Hx, Hy, Hz, Hv). INT8 quantization and a high-performance server ensure real-time operation at 4–5 Hz. In human-robot interaction experiments (90 flights), VLH achieved a 56.7% success rate for target acquisition (mean reach time 21.3 s, pose error 0.24 m) and 100% accuracy in texture discrimination. Generalization tests yielded 70.0% (visual), 54.4% (motion), 40.0% (physical), and 35.0% (semantic) performance on novel tasks. These results demonstrate VLH’s ability to co-evolve haptic feedback with perceptual reasoning and intent, advancing expressive, immersive human-robot interactions.

Keywords: Prosthetics and Exoskeletons
Abstract: Accurate estimation of lower limb motion intention is very important for adaptive control of wearable robot. To address the limited robustness of single sensor mode, this paper proposes a motion prediction framework based on the fusion of surface electromyography (sEMG) and A-mode ultrasound (AUS) signals. The framework uses MS-ALNet deep learning model to process neural driving signals and muscle morphological features through multi-scale feature extraction and attention mechanism. In the walking experiment with 8 subjects, the prediction effects of different signal combinations and mainstream deep learning models were compared. The experimental results show that the prediction accuracy of sEMG+AUS fusion mode is significantly better than that of single mode in hip, knee and ankle joints. The proposed MS-ALNet model is outperforms transformer and CNN-LSTM models in terms of R2 score (both greater than 0.96) and RMSE (ankle joint as low as 1.61 ± 0.2 °). These results suggest that MS-ALNet can effectively use the complementarity of multimodal signals to improve motion intention estimation under the tested conditions. It provides a potential sensing scheme for rehabilitation robots.

Keywords: Health Care Management
Abstract: While wearable sensors have rapidly advanced toward miniaturization and high precision for personalized healthcare, traditional performance gains are reaching a plateau due to a heavy reliance on materials, circuits, and algorithms. This review identifies the exploration of physical mechanisms as the key to overcoming these current bottlenecks. Specifically, we examine three topological physical concepts: exceptional points (EP), which utilize non-Hermitian nonlinear responses to achieve orders-of-magnitude sensitivity enhancements for weak signal detection; spoof surface plasmons (SSP), which leverage engineered structural designs to ensure low-loss signal propagation and solve the interconnect bottleneck on flexible substrates; and valley-Hall effect, which provides robust signal transmission immune to mechanical deformations through topological edge states. By summarizing recent research in wearable sensing and network construction, we provide a physical foundation for the next generation of intelligent, high-performance bioelectronic systems.

Keywords: Industrial and Service Robotics, Collaborative Robots in Manufacturing, Physically Assistive Devices
Abstract: High-intensity industrial labor often places excessive loads on human muscles and joints which significantly increases the risk of developing related musculoskeletal disorders. Industrial exoskeletons are an emerging technology category that is gaining momentum in the manufacturing and construction sectors as they can mitigate physical strain and reduce body wear for workers. The purpose of this article is to provide an up-to-date overview of the various exoskeleton systems that have been developed for specific operational tasks. The article presents the current state of development of these technologies, focusing on the different industrial needs, mechanical structures, and sensing/control strategies. This review uniquely goes beyond the traditional focus on individual joints, analyzing exoskeletons with different joints for specific industrial scenarios, which to meet the support needs of complex industrial tasks. We evaluate the limitations of current assessment methods for industrial exoskeletons. Based on these results, we identify and analyze the design criteria and key performance indicators in manufacturing industrial exoskeletons that can contribute to the large-scale commercial application of advanced industrial exoskeletons.

Keywords: Prosthetics and Exoskeletons, Human Performance Augmentation, Motion Control
Abstract: Active knee exoskeletons should provide effective assistance while remaining lightweight and comfortable to wear. Common knee exoskeletons often place the actuator and structural components on one side of the joint, which can generate undesired interaction loads. Remote transmissions using Bowden cables may also reduce transmission efficiency during gait. To address these issues, this paper presents a lightweight knee exoskeleton that combines a two-side-support layout with a series elastic actuator and cable-roller transmission to reduce undesired interaction loads while improving transmission efficiency and enabling consistent torque control. A time-based gait phase estimation and torque control strategy is developed to provide continuous assistance during level walking and stair ascent. In experiments, the system achieved 98.5% torque tracking accuracy, resistive knee torque below 0.6 N·m in zero-assistance walking, and stable gait-synchronized assistance in both tasks. Preliminary muscle activity results suggested reduced vastus lateralis and vastus medialis activity under assistance, with peak amplitudes decreasing by approximately 5%–12%.

Keywords: Prosthetics and Exoskeletons, Rehabilitation, Physically Assistive Devices
Abstract: This study proposes a soft wearable robotic system that mimics the physiological mechanism of natural respiration, providing noninvasive, bidirectional respiratory assistance for patients with respiratory dysfunction. The results demonstrate that the system enhances peak expiratory flow, improves cough effectiveness and airway clearance, increases breathing amplitude, and improves the accuracy of respiratory rhythm control. Ultrasound evaluation shows increased diaphragm displacement and reduced respiratory rate, resulting in overall improvements in ventilation capacity and breathing efficiency, with strong potential for clinical application.

Keywords: Sensor Fusion, Human Performance Augmentation, Deep Learning in Robotics and Automation
Abstract: Wearable technology has emerged as an important approach for monitoring and supporting human activity in smart industrial environments and daily life assistance applications. Among wearable sensing modalities, inertial measurement units (IMUs) enable real time capture of body motion by measuring acceleration and angular velocity. These sensors are widely used in fitness automation, rehabilitation monitoring, and workplace motion analysis. Automated repetitive action counting (RAC) and repetition interval detection constitute important functions of such systems, supporting personalized exercise feedback and rehabilitation services.

In practice, however, accurately annotating repetition intervals from IMU signals alone is challenging, requiring annotators to consult synchronized video and thereby increasing the labeling burden. Video based models are susceptible to viewpoint variation, occlusion, and background noise, while IMU based models are limited by the lack of semantic context in raw inertial signals.

To address these limitations, this paper proposes a video-IMU multimodal framework for wearable motion analysis. The proposed method aligns representations from both modalities in a shared embedding space via contrastive learning, with the goal of performing RAC and repetition interval detection in an integrated manner. By exploiting the complementary strengths of video and IMU data, the framework aims to reduce the manual annotation burden through the use of video as an auxiliary labeling resource.

Keywords: Industrial and Service Robotics, Motion Control, Robust/Adaptive Control
Abstract: Vectored-thrust cross-domain robots must maintain accurate six-DOF motion under strong nearshore wave disturbances and strict actuator saturation/rate limits. Under such conditions, conventional controller--allocator separation can create a command--execution mismatch: the motion controller evolves based on an ideal generalized force that the constrained allocator cannot fully realize, which may induce windup-like accumulation and degrade 3D stabilization performance. To address this issue, this paper proposes an Allocation-Aware Incremental Nonlinear Dynamic Inversion (AA-INDI) framework that closes the loop with the realized allocated generalized force returned by the allocator. By updating the commanded generalized force around the previously realized input, the proposed architecture preserves command--execution consistency under saturation and vectoring-rate constraints. A fixed-time extended state observer is used to estimate lumped disturbances, a smoothed INDI law generates the incremental motion-control command, and a model predictive control allocator computes feasible thrust-vector commands. Comparative simulations against ADRC on a mission-oriented nearshore aquaculture-cage inspection trajectory under Sea State 2 show that AA-INDI achieves tighter 3D tracking, improved attitude stabilization, and smooth feasible actuation within thrust and vectoring limits.

Keywords: Sensor Fusion, Autonomous Vehicle Navigation, Sensor-based Control
Abstract: Autonomous navigation of unmanned ground vehicles (UGVs) in off-road, GNSS-denied environments relies on the tight coupling of state estimation and local trajectory planning. While receding-horizon optimization planners, notably Model Predictive Control (MPC), have emerged as the standard for dynamically feasible navigation, their recursive feasibility and safety are profoundly predicated on the accuracy, latency, and bandwidth of the underlying state feedback. Furthermore, existing literature predominantly evaluates LiDAR-Inertial Odometry(LIO) algorithms in isolation using open-loop metrics, leaving their system-level impact on online trajectory optimization largely unexplored. To bridge this gap, this paper presents a comprehensive mechanistic analysis of the coupled LIOPlanning system. We systematically investigate how fundamental differences in modern LIO paradigms—namely frame-based, scan-level, and point-by-point updating mechanisms—propagate through the state-feedback loop. By mathematically and empirically analyzing the receding-horizon optimization problem, we demonstrate that delayed or discrete state updates systematically corrupt initial states and spatial constraints, inducing severe trajectory oscillations, redundant replanning, and evasion failures. Conversely, high-bandwidth, point-by-point state updates fundamentally restore initial state fidelity. Extensive real-world off-road experiments validate that point-level LIO provides superior planning consistency and obstacle evasion agility under extreme dynamic conditions, highlighting the critical necessity of holistic, optimization-centric odometry assessments.

Keywords: Intelligent Transportation Systems, Data fusion, Autonomous Vehicle Navigation
Abstract: Cross-domain robots rely on robust Simultaneous Localization and Mapping (SLAM) to operate across heterogeneous environments. However, transitions between different media and locomotion modes introduce distinct vibration characteristics, which can degrade onboard sensor measurements and significantly affect SLAM performance. Despite this challenge, existing studies mainly provide qualitative discussions of vibration effects, while a quantitative understanding remains lacking. To address this problem, we develop a simulation platform for cross-domain robots that enables precise control of vibration characteristics and generates LiDAR and IMU data under different vibration conditions. Based on this platform, we construct the Vibration LiDAR-IMU (Vib-LI) Dataset, investigating the impact of vibration frequency and vibration amplitude on SLAM algorithms and analyzing the reasons behind the performance degradation caused by vibrations. We believe this research provides valuable insights for vibration reduction in cross-domain robots and the design of SLAM algorithms.

Keywords: Manipulation Planning, Motion and Path Planning, Collision Avoidance
Abstract: Installing nozzle dams in nuclear power plant steam generators is a high-radiation maintenance task that is still largely performed manually by human workers. Automating this process with robot can reduce worker exposure, however, the challenge lies in guiding a heavy dam segment inside the extremely confined space, where conventional path-planning methods can struggle to satisfy both feasibility and safety-clearance requirements. In this paper, we present a deep reinforcement learning approach for autonomous navigation and continuous control of an 8-DoF redundant manipulator designed specifically for nozzle dam installation. The task is formulated as a Markov decision process and solved using Proximal Policy Optimization (PPO). Training is carried out in 1,024 parallel simulation environments with a multi-objective reward function that balances target pose tracking, spatial safety margins, and mechanical smoothness. Simulation results show that the trained policy generates smooth joint trajectories while maintaining larger clearances from obstacles, without requiring manual calibration. These results indicate that the proposed method is well suited to steam generator nozzle dam installation in the confined environment, thereby improving the safety of this task.

Keywords: Collaborative Robots in Manufacturing, Factory Automation, Intelligent and Flexible Manufacturing
Abstract: 以提升协作操作效率 自动仓储中的升降机和四向穿梭机 在灵活生产场景下的系统，一个双级 协作优化方法，集成改进后的 提出了A*算法和模拟退火。在 在较低层次，引入时间参数t于 传统的A*算法。通过确定占用情况 节点在时间 t 时节点的状态，潜在冲突 可以提前识别并避免，从而实现 冲突感知的最短可行路径规划 排班。同时，出站任务被分为以下几类 三个优先级，分别是高、中、低， 这些优先级仅作为 节点冲突场景中的路权分配。在 上层建立任务分配模型，通过 目标是最小化总任务完成时间， 并采用模拟退

Keywords: Telerobotics and Teleoperation, AI-Based Methods, Human Factors and Human-in-the-Loop
Abstract: In-contact teleoperation tasks, such as surface polishing and fragile object manipulation, require coordinated control of motion and interaction force as well as timely adjustments from the human operator. This paper presents a wearable multimodal teleoperation framework that integrates a data-driven wearable control interface with vibrotactile biofeedback. Using a single commercial armband, the proposed interface estimates both the operator’s hand motion and intended interaction force from inertial measurements and surface electromyography (sEMG) signals, enabling the generation of end-effector pose and force commands simultaneously. To assist the operator during task execution, a vibrotactile device delivers biofeedback that encodes the inferred force command, allowing the user to regulate interaction force even before physical contact occurs. The framework also supports intuitive switching between Cartesian motion control and hybrid motion–force control for different task stages. A user study involving two representative teleoperation tasks—surface polishing and fragile object pick-and-place—verifies the effectiveness of the proposed robot control framework.

Keywords: Telerobotics and Teleoperation, Deep Learning in Robotics and Automation, Agricultural Automation
Abstract: Teleoperation enables human supervision and control of agricultural mobile robots in outdoor field environments. However, limitations in the perception–action loop caused by network infrastructure hinder the reliability of video feeds delivered to teleoperators. These limitations introduce variable delays in visual feedback, causing the received images to significantly differ from the robot’s current viewpoint. This paper proposes an extended-horizon frame prediction framework based on a new delay-informed state predictor to generate delay-compensated video feeds for teleoperators. The proposed framework integrates delay-aware state prediction with geometry-based viewpoint synthesis to reconstruct the robot’s current visual observation from delayed inputs. By utilizing estimated delays and delayed camera state information, the framework effectively generates motion-constrained future frames that align with the robot’s current viewpoint. Extensive evaluations in a real-time simulation setup built from real-world agricultural robot datasets demonstrate that the proposed approach generates more accurate images and improves visual alignment compared to existing methods, even under large and highly variable delays. These results indicate that the framework could positively impact real-time teleoperation of agricultural robots in environments with dense crop rows.

Keywords: Agricultural Automation, Computer Vision in Automation, Reactive and Sensor-Based Planning
Abstract: Broadcast seeding is widely adopted for short-cycle leafy vegetable cultivation because of its low labor and planting costs. However, the resulting irregular crop distribution and dense plant clusters create significant challenges for automated thinning and weeding operations. This paper presents an integrated autonomous robotic system for broadcast-seeded crop management that combines pose-aware crop perception, a Vigor-Guided Thinning Algorithm (VGTA), robotic manipulation, and crop distribution mapping. The proposed perception framework employs pose-aware detection to improve crop center localization under dense occlusion conditions. To support thinning decision-making, VGTA integrates plant vigor estimation with KD-tree-based spatial indexing, enabling the selection of vigorous seedlings while satisfying spacing constraints. Plant vigor is quantified using image-derived plant area extracted from RGB images. Experiments conducted in a greenhouse environment demonstrated that the proposed VGTA increased the average retained crop area by approximately 45% compared with a conventional distance-based thinning strategy. The proposed algorithm achieved real-time performance, with an average computation time of 0.56 ms per working window and a computational complexity of O(n log n). In addition, the system generates crop distribution maps that support future precision agriculture applications. The results demonstrate the effectiveness of the proposed framework for autonomous thinning and weeding in broadcast-seeded vegetable cultivation.

Keywords: Task Planning, Industrial and Service Robotics, Manipulation Planning
Abstract: We present a fast grasp planner for the Grasp EveryThing (GET) asymmetrical gripper. GET-2D-1.0 uses a single-view RGB-D image, the Ferrari-Canny metric, and a novel sampling strategy. We also present GET-3D-1.0, a mesh-based method using a 3D gripper model and ray-tracing. We evaluate both grasp planners against baselines with physical experiments, which suggest that GET-2D-1.0 can improve over a bounding box baseline by over 40% in lift success, shake survival, and force resistance. Experiments with GET-3D-1.0 suggest slight improvement compared to GET-2D-1.0 on lift success and shake survival, but are more computationally expensive, averaging 17 seconds of planning compared to 683 ms for GET-2D-1.0. Files and fabrication instructions are available on our project site and GitHub for GET jaws for the Robotiq 2F-85 gripper and the standard YAM robot. https://apgoldberg1.github.io/GET_Planning

Keywords: Optimization and Optimal Control, Cyber-physical Production Systems and Industry 4.0, Modelling, Simulation and Validation of Cyber-physical Energy Systems
Abstract: Real-time optimal control of complex hybrid systems—governed by both continuous thermodynamic dynamics and discrete topological switching—presents a formidable Mixed-Integer Nonlinear Programming (MINLP) challenge. Traditional exact solvers often experience severe performance degradation, or ``engineering suffocation,'' under strict computational time budgets and tightened robust safety constraints. To address this critical bottleneck, we propose the Quantum-Classical Collaborative MINLP Dynamic Optimization (QC-MINLP-DO) framework, which integrates the Quantum Approximate Optimization Algorithm (QAOA) with robust mode-dependent Tube-MPC. By enforcing a strict dual-timescale separation, the framework isolates continuous worst-case safety from discrete combinatorial exploration. A fast-timescale Tube-MPC guarantees absolute recursive feasibility and physical safety against bounded disturbances. Asynchronously, a cost-aware meta-scheduler maps a trust-region-bounded pseudo-Boolean surrogate to a sparse Ising Hamiltonian, utilizing the QAOA backend to orthogonally sample high-quality topological candidates. Crucially, a two-tiered Cross-Domain Admissibility Veto Protocol strictly intercepts any unsafe or infeasible proposals prior to execution. Extensive time-domain simulations on a green-hydrogen microgrid demonstrate the framework's efficacy across varying Feasibility Fragmentation Index (FI) regimes. Remarkably, under a strict 1.5-second Time-to-Solution (TTS) budget and extreme FI penalties, the quantum-assisted framework completely avoids the choking phenomenon of classical solvers, delivering up to a 15.3x reduction in Non-Update Probability (NUP) while mathematically guaranteeing zero physical constraint violations.

Keywords: Machine learning, Failure Detection and Recovery
Abstract: Fault signals of rolling bearings in wind turbine drivetrain systems often exhibit characteristics such as nonlinearity, non-stationarity, and low signal-to-noise ratio. Under complex operating conditions, traditional signal processing methods and shallow classification models struggle to maintain high recognition stability and diagnostic accuracy. To address this issue, this paper proposes a fault diagnosis method that combines Variational Mode Decomposition with a deep convolutional bidirectional long short-term memory network. After data augmentation is applied to the original vibration signals, VMD is used to decompose the samples into multiple intrinsic mode functions, which are then used to construct multi-channel temporal inputs. Furthermore, a VMD-CNN-SE-BiLSTM network is designed, in which multi-layer one-dimensional convolutions are employed to extract local fault features, the SE attention mechanism is introduced for channel recalibration, and BiLSTM is incorporated to capture temporal dependencies in deep features. Finally, a global average pooling layer over time and a fully connected layer are used to classify fault categories. Experimental results on the Case Western Reserve University dataset show that the proposed method achieves a recognition accuracy of 99.4%. By integrating modal decomposition, local feature representation, and temporal dependency modeling, the proposed method provides an effective solution for rolling bearing fault diagnosis under complex operating conditions.

Keywords: Manufacturing, Maintenance and Supply Chains, Factory Automation, Intelligent and Flexible Manufacturing
Abstract: False alarms remain a primary obstacle to the industrial scalability of predictive maintenance (PdM) deployment. While existing research focuses on refining Remaining Useful Life (RUL) prediction accuracy, relatively little attention has been paid to the systemic governance required to distinguish true degradation from exogenous disturbances, sensor drifts, or transient spikes. This paper proposes SENTRYPdM, a unified and testable architecture designed to augment standard PdM pipelines with principled alarm governance. The architecture integrates several core mechanisms, including: (i) signal hygiene and exogenous debiasing, which use intervention-aware forecasting to filter operational offsets; (ii) cross-horizon agreement (CHA) to enforce RUL consistency across multiple look-ahead windows; (iii) residual change-point diagnostics to differentiate structural failures from reversible excursions; (iv) surge disambiguation for catastrophic event detection using extreme-value tails; and (v) adaptive decision logic (ADL++) to synthesize multi-modal evidence into cost-optimized alerts. Validation using BOSCH DRIE production data and a stress test across twelve diverse failure scenarios (Cases A–L) demonstrates that SENTRYPdM significantly reduces the mean Beginning-to-End (BTE) error from 1.686 to 0.239, without sacrificing early-warning lead time. The results indicate that the proposed framework can function as a “plug-and-play” layer to enhance operator trust and operational reliability in smart manufacturing environments.

Keywords: Model Learning for Control, Deep Learning in Robotics and Automation, Task Planning
Abstract: Vision-language-action (VLA) models have recently emerged as a promising paradigm for robotic control, enabling end-to-end policies that ground natural language instructions into visuomotor actions. However, current VLAs often struggle to satisfy precise task constraints, such as stopping based on numeric thresholds, since their observation-to-action mappings are implicitly shaped by training data and lack explicit mechanisms for condition monitoring. In this work, we propose CLAW (CLIP-Language-Action for Weight), a framework that decouples condition evaluation from action generation. CLAW leverages a fine-tuned CLIP model as a lightweight prompt generator, which continuously monitors the digital readout of a scale and produces discrete directives based on task-specific weight thresholds. These prompts are then consumed by pi_0, a flow-based VLA policy, which integrates the prompts with multi-view camera observations to produce continuous robot actions. This design enables CLAW to combine symbolic weight reasoning with high-frequency visuomotor control. We validate CLAW on three experimental setups: single-object grasping and mixed-object tasks requiring dual-arm manipulation. Across all conditions, CLAW achieves a 100% success rate, while the fine-tuned pi_0 reaches at most 35%. href{https://youtu.be/MuMYj2QgReI}{A video} of our paper is available online.

Keywords: Motion and Path Planning, Control Architectures and Programming, Motion Control
Abstract: The specialized tilt-rotor mechanism of Uncrewed Aerial Vehicles (UAVs) facilitates a flexible transition between multi-rotor and fixed-wing modes through an intermediate transitional phase, significantly expanding the aircraft's operational envelope and maneuverability.While conventional studies characterize the transitional state as a transient and suboptimal phase, its inherent potential for high agile maneuvers is often underestimated.This paper proposes a transitional flight methodology based on continuous thrust vectoring control.By precisely modulating the motor tilt angles, the proposed approach achieves synergistic control of the UAV’s attitude and thrust vectoring, enabling the execution of agile aerobatic maneuvers during the transition period.Comparative numerical simulations demonstrate that, although multi-rotor mode can perform similar tasks, utilizing the transitional state yields superior control fidelity, accelerated dynamic response, and enhanced trajectory tracking performance.The results validate that leveraging transitional flight for agile aerobatic maneuvers offers distinct performance advantages and substantial potential for advanced aerial applications.

Keywords: Motion and Path Planning, Motion Control
Abstract: Collaborative multi-agent object transportation with mobile robots is becoming very popular due to its large potential applications in various fields, from construction, smart warehousing, mining, to space colonization. In multi-agent scenarios, agents need to make rational decisions that influence the behaviour of other agents to act effectively for achieving a common goal. In this work, we explore the application of game theoretic methods for solving the collaborative object transportation problem. Our work considers non-holonomic mobile manipulators carrying a rigid object from an initial position to a predefined goal position in an obstacle-laden environment. The proposed work is based on the computation of local Stackelberg Equilibrium, which assumes a predefined hierarchy within the agents. In comparison to the Centralized Model Predictive Control based method our approach shows higher task completion rate. We discuss the effect of predefined role assignment of the agents and argue that a dynamic role selection can be beneficial. Numerical simulations with two and more agents show the extensibility of our framework to multiple agents. Further, experiments on hardware indicate favorable performance that corroborates the effectiveness of the theory.

Keywords: Sensor-based Control, Computer Vision in Automation, Agent-Based Systems
Abstract: Orientation-aware manipulation is essential in post-harvest agricultural processing, where produce must be grasped and placed in consistent configurations. This paper presents ROG-Grasp, a geometry-based robotic grasping and placement framework that estimates the produce orientation from root surface geometry using RGB-D perception. A YOLO-based root detector and point cloud plane fitting are used to infer the root normal, enabling stable grasp pose generation and orientation-constrained Cartesian motion planning. Experiments on tomatoes and onions demonstrate high success rates and stable execution time in both isolated and cluttered scenarios. Compared with vision–language–action (VLA) policies, the proposed method achieves more reliable and accurate grasp completion with faster execution. These results highlight the effectiveness of geometry-driven perception for practical orientation-controlled manipulation tasks.

Keywords: Sensorimotor Learning, Physical Human-Robot Interaction, Force and Tactile Sensing
Abstract: Estimation of robotic dynamic states in physical human-robot interaction (pHRI) is crucial for robots to handle real-world uncertainties. Since wearable tactile sensor arrays have been increasingly applied in robots, the tactile sensory outcomes (TSOs) during pHRIs would provide practical and ideal information about dynamic states of a robot as responses of its actions. In this work, an

embodied learning strategy is introduced to promote the formation of a deep learning-based embodied haptic model (DL-EH), enabling robots to implicitly learn their tactile sensorimotor dynamics in pHRI. Functionally, DL-EH adopts a spatial-temporal assist module (embodied haptic embedding) to assist a compact forward predictor (multimodal mapping and sensory prediction) in learning TSO prediction. Then, it can realize an end-to-end real-time TSO prediction of robot actions in pHRI based on current tactile signals without historical sequences. In real-world experiments, the effectiveness of DL-EH is demonstrated, especially in long-term TSO predictions. The generalization test is performed to validate DL-EH's generality in new pHRI scenarios. Finally, a robotic framework via the proposed method is proposed for proof-of-concept in the pHRI application.

Keywords: Service Robotics, Human-Robot Collaboration, Social HRI
Abstract: Shopping service robot (SSR) is designed to offer a superior shopping experience through its continuous target following and companion services. However, the SSR encounters difficulties in complex environments, including visual occlusions and spatial constraints in narrow aisles. To address these difficulties, the study proposes a shopping service robot framework that integrates innovative perception, following control, and path planning modules. Firstly, the perception module employs a multi-sensor fusion method for human target tracking, integrating both red-green-blue-depth (RGB-D) camera and wireless electromagnetic (WEM) data by using an extended Kalman filter (EKF). Secondly, the target-following control module employs an omnidirectional constrained control law, which ensures synchronized orientation alignment between the SSR and the human target. Finally, the path planning module employs topological mapping to encode intersection geometries as path nodes, thereby guiding the SSR pass through narrow shelf aisles. In real supermarkets, we evaluated the target tracking approach under shelf occlusions and the human following task within narrow aisles. Experimental results demonstrate that the visual-WEM tracking approach achieves a pose tracking accuracy of (4.56 mm, 2.98°) under shelf occlusions. This study establishes the feasibility of human-robot collaboration in facilitating a hands-free shopping experience, highlighting its potential as a substitute for conventional shopping carts.

Keywords: Surveillance Systems, Data fusion, Probability and Statistical Methods
Abstract: Active sonar target detection in shallow water is severely degraded by strong reverberation and multipath interference. While conventional beamforming (CBF) provides a computationally efficient preprocessing front-end for real-time systems, it still suffers from sidelobe leakage and limited reverberation suppression capability. Furthermore, traditional low-rank and sparse decomposition methods, such as Robust Principal Component Analysis (RPCA), often fail to distinguish true targets from non-stationary reverberation due to a lack of physical acoustic priors. To address these challenges, this paper proposes a real-time, range-dependent hierarchical multi-feature structure-aware RPCA framework. Following CBF preprocessing, the proposed architecture employs a tensor-based singular value decomposition to suppress near-field direct-wave interference. For the far-field region, an adaptive Alternating Direction Method of Multipliers (ADMM) scheme is developed. It dynamically fuses three physical constraints: spectro-temporal linear frequency modulation (LFM) structure, spatial beam diffusion, and local statistical stability.This multi-feature soft-masking mechanism guides the sparse optimization process by favoring physically meaningful target echoes while suppressing background reverberation. Experimental results demonstrate that the proposed framework significantly improves the signal-to-reverberation ratio and detection probability compared to baseline methods, while maintaining a low computational complexity highly suitable for autonomous underwater vehicles.

Keywords: Task Planning, Deep Learning in Robotics and Automation
Abstract: Vision-Language-Action (VLA) models learn generalist robot manipulation policies by mapping language instructions and visual observations to continuous actions through imitation learning. However, their performance degrades on long-horizon tasks, particularly when sub-tasks admit multiple valid execution orders. Since different demonstrators naturally choose different permutations, the same visual observation is paired with actions for different sub-tasks across the training set. This one-to-many ambiguity introduces contradictory supervision signals, causing the learned policy to oscillate among sub-tasks at inference time. To address this, we propose Stage-Aware VLA, which conditions the policy on an explicit progress state that tracks sub-task completion at runtime, disambiguating the observation-action mapping. Specifically, we introduce a learnable stage query that is jointly processed with visual and language inputs by the VLM to predict the currently active sub-task. Based on this prediction, a stage manager updates the progress state accordingly to guide the policy through the task stages. Experiments on a desk organization task with four sub-tasks show that Stage-Aware VLA improves overall success rate and yields more balanced sub-task completion.

Keywords: Failure Detection and Recovery, Task Planning, Causal Models
Abstract: To achieve high levels of autonomy, robotic systems must be able to handle unknown anomalies that occur during execution. Unknown anomalies have never been seen by the robot or system and so no known recovery behavior exists. Handling anomalies requires three steps: detecting the anomaly, identifying its root cause, and recovering from it. In this paper we present CiTAMP (causal intervention task and motion planning), which addresses unknown anomaly identification and recovery. Causal interventions provide a way of answering `what-if?' queries about the system. Given a detected anomaly, we use parallelized Monte Carlo tree search (MCTS) to try causal interventions in a simulator that does not contain the robot. From this we obtain a state from which nominal execution can continue. We then pass this state to an existing task and motion planner to synthesize a sequence of kinematically feasible recovery actions for the robot. We demonstrate the efficacy of our approach on four simulated block manipulation tasks.

Keywords: Failure Detection and Recovery, Task Planning, Probability and Statistical Methods
Abstract: Autonomous robotic systems are susceptible to anomalies, i.e., deviations from expected behavior. Anomalies are caused by environmental changes, execution errors, or stochastic action outcomes, which can prevent task completion. If unaddressed, these anomalies can compromise the robot’s ability to complete its tasks. Existing anomaly-handling approaches, namely proactive and reactive methods, fall short in effectively addressing failures that have already occurred. Such failures require structured diagnosis and deliberation rather than immediate, predefined responses. We refer to this as post-hoc anomaly handling. To address this, we propose BN-TAMP, a Bayesian Network and Task and Motion Planning (TAMP) framework for post-hoc anomaly handling. BN-TAMP detects anomalies via rule-based threshold checks on symbolic state representations, identifies them through a learned Bayesian Network to infer the feasible continuous goal state, and generates corrective actions using PDDLStream to transition the system back to an anomaly-free state. Our evaluation considers two validation regimes, stepwise (after each action) and final (only at the end). Results demonstrate that BN-TAMP achieves higher success rates with fewer actions and shorter execution times, highlighting its effectiveness in post-hoc anomaly handling.

Keywords: Force and Tactile Sensing, Haptics and Haptic Interfaces, Sensor-based Control
Abstract: Tactile sensing is essential for humanoid robots to interact safely with humans and operate in complex environments. However, existing tactile sensing systems often suffer from limited sensing modalities and insufficient scalability for large-area deployment on robotic surfaces. This paper presents a scalable dual-mode tactile sensing system composed of seven hexagonal modules. Each sensing module includes six pressure-sensing regions and one central capacitive proximity-sensing region, enabling continuous perception from object approach to physical contact. The pressure and proximity sensing functions are realized through piezoresistive transduction and interdigitated electrodes, respectively. By combining flexible PDMS encapsulation with arched inter-module interconnects, the system improves conformal scalability on curved robotic surfaces. Additionally, an event-triggered serial communication method is employed to reduce unnecessary signal transmission, while redundant interconnections are introduced to enhance system robustness. As a result, the system achieves a stable data update rate above 50 Hz. Experimental results demonstrate a stable pressure sensing range of 0-8.3 N and a proximity detection distance of up to 30.0 mm, with maximum sensitivities of 2.542 V/N and 0.248 pF/mm for pressure and proximity sensing, respectively. Application experiments further verify the system’s capability to detect approaching objects and reconstruct tactile information distributions, highlighting its potential for scalable robotic tactile perception and safe human-robot interaction.

Keywords: Force and Tactile Sensing, Machine learning, Sensor Fusion
Abstract: In the field of robotic manipulation, accurately estimating the pose of an object within a multi-fingered dexterous hand—known as in-hand pose estimation—is essential for achieving fine and dexterous manipulation. However, most existing in-hand pose estimation methods rely on two-finger grippers and external vision, which are not well-suited to the challenges posed by multi-fingered dexterous hands, where severe occlusions and higher object mobility complicate the problem. To address this issue, we propose an approach that relies solely on vision-based tactile sensors mounted on the fingertips of the dexterous hand, together with a tactile array sensor embedded in the palm. By fusing contact data from these two sensing modalities, our method estimates the in-hand object pose without external vision. To our knowledge, this may be the first attempt to combine vision-based tactile sensors and a palm tactile array for in-hand object pose estimation. Furthermore, we develop a simulation approach on the Isaac Sim platform that can generate synthetic data from vision-based tactile sensors in a large-scale, parallel, and rapid manner, enabling us to train our network entirely with simulated data and providing a valuable foundation for future reinforcement learning. Both simulation and real-world experiments validate the effectiveness of the proposed approach.

Keywords: Foundations of Automation, Automation in Construction, Industrial and Service Robotics
Abstract: This paper describes a multi-robot solution for pick-and-place operations, that creates and exploits mechanical constraints to control the payload through interactions. The task requirements are split across a strong, but imprecise, crane and an accurate, but weak, dual-arm manipulator. First, the design of the manipulator is discussed. Its task is to align its end-effectors with features in the environment and create additional mechanical constraints. Second, a coordinated control strategy between agents is described in order to accurately position and align a payload relative to another object in the environment. The crane controller exploits interactions between the payload and the mechanical constraints created by the manipulator to control the payload and reduce its state uncertainty.

A box stacking use case was performed to validate the method. The task could be successfully executed, even when tracking errors were superimposed on the position- and velocity setpoints of the crane control. This highlights the robustness of the approach, and the possibility to implement it on existing, imprecise cranes with low control bandwidth.

Keywords: Medical Robots and Systems, Tendon/Wire Mechanism
Abstract: Bladder cancer is commonly treated by transurethral resection, but conventional rigid instruments provide limited distal dexterity and restrict access to anatomically complex intravesical regions. This paper presents a 5-DoF dual-segment continuum waterjet robot for transurethral bladder tumor ablation. The system integrates a high-pressure cavitating waterjet nozzle at the distal end of a 6-mm flexible arm, enabling independent bending and orientation control. A kinematic model and model-based tendon compensation method are developed to reduce proximal-distal coupling, and a master-slave control architecture is implemented for operator-commanded teleoperation. Workspace simulation, tendon coupling compensation tests, positioning repeatability measurements, and multi-site targeting experiments in a bladder phantom demonstrate the feasibility of flexible continuum manipulation combined with waterjet ablation for minimally invasive intravesical surgery.

Keywords: Modelling, Simulation and Optimization in Healthcare, Discrete Event Dynamic Automation Systems
Abstract: This paper addresses the evaluation of energy expenditure associated with outdoor walking for elderlies. We propose a Markov decision process (MDP), an original framework that models walking as a sequential decision problem. This framework allows us to evaluate any walking strategy, determine the optimal strategy that minimizes a cost function subject to energy reserve constraints, and analyze their properties. Numerical experiments further demonstrate how these strategies adapt to different cost functions or to the preferences of older adults. Since the framework relies on simplifying assumptions, this contribution is primarily methodological and lays the foundation for a digital twin capable of proposing routes adapted to the physiological capabilities of seniors.

Keywords: Modelling, Simulation and Optimization in Healthcare, Health Care Management, Scheduling in Healthcare
Abstract: As an advanced radiation treatment, proton therapy is a highly targeted treatment for tumors and can minimize damage to surrounding normal tissues. A typical proton therapy system consists of one proton accelerator and multiple gantries. At any time, the proton accelerator can only serve one gantry. Such a resource-sharing feature of the proton therapy delivery makes patients wait on the treatment couches of the gantries. Long waiting time compromises treatment precision, causes patient muscle injuries, and reduces the delivery efficiency. To address this challenge, we propose a two-stage stochastic programming model for real-time control of proton therapy delivery. The model considers uncertainties of durations in each step of proton therapy delivery, minimizes patient waiting times, and maximizes system utilization. In addition, we provide an aggregation approach and a real-time maintained model to improve the computational efficiency of stochastic programming, enabling it to handle this real-time control problem. Numerical experiments show an improved performance of the proposed method. For the metrics of average beam waiting time and total service time, the two-stage stochastic programming model achieves improvements of 62% and 5% over the naive method, and 49% and 10% over the simulation optimization gantry call-back control (SOGCC) method, respectively. For average decision time, the proposed aggregation and real-time maintained model improve computational efficiency by 25%. Overall, by using a two-stage stochastic programming model to a real-time control problem of proton therapy delivery, we obtain fast and high-quality solutions.

Keywords: Modelling, Simulation and Validation of Cyber-physical Energy Systems, Power and Energy Systems automation, Sustainability and Green Automation
Abstract: High renewable energy penetration brings urgent accommodation challenges to power systems. Interconnected multi-regional integrated energy systems (MRIES) offer an effective solution via cross-energy and cross-spatial balancing. To fully exploit the inherent storage potential of the interconnected district heating network (DHN) within MRIES, this paper proposes a coordinated optimization framework for electricity-heat-hydrogen coupled MRIES. First, a unified scheduling model is established, which fully incorporates the operation constraints of multi-energy stations, cross-regional multi-energy interaction mechanisms, and the refined hydraulic and dynamic thermal models of DHN, to accurately characterize the inter-temporal thermal storage flexibility from pipeline heat transport delay and thermal inertia. To address the non-convex bilinear terms introduced by variable flow and variable temperature (VF-VT) regulation, the McCormick envelope is adopted to realize convex relaxation, and a sequential bound tightening (SBT) algorithm is further developed to iteratively narrow the feasible range of variables, enhance the tightness of relaxation. Numerical results demonstrate that the proposed method can effectively leverage the flexibility of heating networks, and achieve significant advantages in both operational economy and renewable energy accommodation.

Keywords: Power and Energy Systems automation, Machine learning
Abstract: User-side peer-to-peer (P2P) multi-energy trading involves a closed-loop interaction between local energy management and market operation, as users' decisions jointly determine clearing outcomes, which in turn shape subsequent decisions. This paper develops a user-side P2P multi-energy trading framework covering electricity, gas, and heat, in which users trade directly with one another while the external market serves only as a non-strategic price reference and fallback channel. User decision making is formulated as a parameter-shared multi-agent reinforcement learning problem, with a market-clearing mechanism explicitly embedded in the environment dynamics. Resulting clearing prices, traded volumes, and matching outcomes are fed into subsequent user observations. A PPO-based method is employed to learn decentralized energy-management and trading policies. Simulation results in a five-user proof-of-concept case demonstrate convergent training, clearing prices within prescribed external-market bounds, and heterogeneous inter-user trading relationships, supporting the feasibility of integrating market clearing and multi-agent policy learning for closed-loop user-side P2P multi-energy trading.

Keywords: Motion Control, Autonomous Vehicle Navigation
Abstract: To address the fixed-point hovering problem of a Remotely Operated Vehicle (ROV) near the free surface under significant wave conditions, this paper proposes a disturbance-rejection control framework that integrates wave-disturbance preview with reinforcement learning. Exploiting the deterministic propagation and dispersion properties of irregular waves, the Deterministic Sea Wave Prediction (DSWP) is employed to infer a future disturbance sequence at the ROV location from upstream wave-elevation measurements. Furthermore, using group-velocity bounds of the principal energy band in the wave spectrum, an Effective Preview Window (EPW) is deterministically computed to retain physically credible segments of the prediction and discard unreliable long-horizon parts distorted by dispersion. To avoid the dimensionality explosion and training instability caused by directly concatenating long disturbance sequences, an in-window exponentially weighted aggregation and a low-dimensional feature compression mechanism are designed, converting the future disturbance preview into an interpretable, low-dimensional input that enhances the policy's feedforward disturbance-rejection capability. A 2D hovering scenario (surge--heave--pitch) is built with JONSWAP-spectrum-driven waves and Morison-equation-based load modeling, and the proposed method is systematically compared against no-preview RL, full-sequence-preview RL, and a PID baseline across multiple sea states (Hs=0.20 and 0.40 m). The results show that the EPW-based compressed preview significantly improves hovering accuracy and robustness.

Keywords: Motion Control, Motion and Path Planning, Optimization and Optimal Control
Abstract: Bimanual manipulation is a fundamental capability for humanoid robots operating in real-world environments, however it remains challenging due to coupled workspaces, closed-chain constraints, and singularity-prone kinematics. This paper proposes Operational Space Bimanual Manipulator Controller (OSBMC), an optimization-based operational space control framework for bimanual manipulators that explicitly accounts for coupled task execution and singularity avoidance. Inspired by control barrier function (CBF) formulations, the proposed method integrates task-space tracking objectives with singularity-aware constraints in a unified quadratic programming framework. Unlike conventional decoupled approaches that easily violate grasp constraints near singularities, OSBMC regulates the dual-arm system to preserve strict closed-chain consistency. The controller is implemented on a 14-DOF humanoid upper body, and its effectiveness is validated through both simulation and a real-world bimanual grinding task. Experimental results demonstrate that the proposed approach actively avoids singular configurations while strictly preserving closed-chain consistency, enabling robust and highly practical dual-arm manipulation.

Keywords: Optimization and Optimal Control, Agricultural Automation, Planning, Scheduling and Coordination
Abstract: Heterogeneous unmanned aerial vehicle (UAV) coordination in precision agriculture is increasingly challenged by severe bandwidth constraints and limited computational resources. Existing distributed algorithms often suffer from prohibitive signaling overhead in dynamic environments, while event-triggered mechanisms in discrete strategy spaces are prone to instability and decision oscillations. To address these issues, this paper proposes the Event-Triggered Best Response with Inertia (ET-BRI) algorithm. We model multi-UAV interactions as a congestion game and establish its equivalence to an exact potential game (EPG), providing a rigorous theoretical foundation for coordination. Within this framework, ET-BRI utilizes a deterministic inertia threshold lambda to filter marginal utility gains, effectively pruning redundant signaling and dampening oscillations. Extensive simulations over 100 trials involving dynamic tasks show that ET-BRI achieves a robust completion count of 47.3 +/- 1.1, matching the performance of complex D-MPC and TVCBLL algorithms. Notably, it maintains a low computational latency (4.85 ms) while reducing communication frequency by over 98%, validating ET-BRI as a scalable and resource-efficient solution for bandwidth-constrained field operations.

Keywords: Optimization and Optimal Control, Smart Grids, Reinforcement
Abstract: This paper develops a distributed reinforcement learning (RL) method to coordinate cooperative microgrids (MGs). The high uncertainty of power loads and renewable energy sources motivate the operator to perform real-time dispatch. On the one hand, the existing online methods usually utilize approximate models that result in intractable constraint violation. A common method is to relax it as a chance constraint, while it is still hard to ensure its satisfaction in practice. On the other hand, some MGs may hope to preserve the private information on their local costs and states. To address these problems, we make the following contributions. First, the coordination problem is reformulated as a constrained multi-agent Markov decision process. Second, the distributed RL algorithm with a theoretical convergence guarantee is developed. Third, to further preserve the local private information and improve the performance, this algorithm is modified by adding a local feature extraction module for each agent. This module could also be regarded as an encryption module for the local state information. Fourth, numerical experiments are carried out to validate the effectiveness of the modified algorithm. Note to Practitioners—In multi-MG systems, voltage security is significant to energy reliability. However, the MGs are controlled by different owners who are concerned about privacy. The conventional model-based methods require detailed information to construct the closed model. And the inaccurate model induces intractable abnormal voltage problems. To address these issues, we propose to restrict the off-limit frequencies by chance constraint since the low-frequency off-limits could be adjusted by tertiary control. Additionally, we develop a distributed RL algorithm to optimize the c

Keywords: Optimization and Optimal Control, Task Planning, Motion and Path Planning
Abstract: In mobile welding automation, a mobile platform is required to carry a welding manipulator while maintaining positional accuracy and stability. The elevated center of mass introduced by the manipulator increases the system's sensitivity to base roll and pitch variations, even under small disturbances. This sensitivity becomes particularly problematic when traversing repeated T-bar structures, where abrupt transitions in support conditions induce pronounced attitude changes. Therefore, to secure stability during T-bar traversal, this paper proposes a quadruped-wheeled robot system that maintains base levelness while ensuring maneuverability.

The proposed platform combines omnidirectional driving using Mecanum wheels at the leg ends with a sliding-electromagnet module beneath the base. During traversal, the electromagnets adhere to the T-bar top surface to form an additional support point, and the sliding actuation constrains base advancement to linear motion, mechanically suppressing roll and pitch variations. A finite-state-machine-based traversal sequence coordinates electromagnet engagement and release, sliding actuation, and leg reconfiguration. Experiments show that base roll and pitch remain within 4 degrees during the traversal segment, and power spectral density analysis of acceleration at a representative point indicates that the 0 to 5 Hz low-frequency band is dominant, meaning that vibrations are governed more by base attitude changes and body motion than by high-frequency disturbances induced by T-bar contact.

Keywords: Planning, Scheduling and Coordination, Reinforcement, Collision Avoidance
Abstract: Waste cranes are the core hubs connecting waste pits and incinerators in waste-to-energy plants, responsible for handling tasks such as waste feeding, turning and transfer. Due to the dynamic arrival of garbage trucks and the fluctuating feeding demands of incinerators, real-time scheduling of waste cranes to minimize task waiting time poses a significant challenge. Traditional heuristic rules lack long-term foresight, while exact methods struggle with computation time. This paper proposes a deep reinforcement learning model based on graph neural networks. In the model, we extend the Extended Temporal Distance (ETD) method to address crane collision avoidance specifically in waste crane scenarios. Besides, tailored to the problem's characteristics, we design a heterogeneous graph to represent the states of the cranes and tasks. Finally, the model is trained using the proximal policy optimization. Preliminary results demonstrate that our approach outperforms conventional heuristic rules and, due to its long-term perspective, even surpasses short-sighted rolling horizon method based on exact algorithm.

Keywords: Planning, Scheduling and Coordination, Reinforcement, Intelligent and Flexible Manufacturing
Abstract: Efficient job-shop scheduling with transportation resources is critical for high-performance manufacturing. With the rise of "decentralized factories", multi-agent reinforcement learning has emerged as a promising approach for the combined scheduling of production and transportation tasks. Prior work has largely focused on developing novel cooperative architectures while overlooking the question of when joint training is necessary. Joint training denotes the simultaneous training of job and automatic guided vehicle scheduling agents, whereas modular training involves independently training each agent followed by post-hoc integration. In this study, we systematically investigate the conditions under which joint training is essential for optimal performance in the job-shop scheduling problem with transportation resources. Through a rigorous sensitivity analysis of resource scarcity and temporal dominance, we quantify the coordination gap --- the performance difference between these two training modalities. In our evaluation, joint training outperforms the majority of dispatching rule combinations and modular training approaches. However, the coordination gap advantage diminishes in bottleneck environments, particularly under severe transport and processing constraints. These findings indicate that modular training represents a viable alternative in environments where a single scheduling task dominates. Overall, our work provides practical guidance for selecting between training modalities based on environmental conditions, enabling decision-makers to optimize reinforcement learning-based scheduling performance.

Keywords: Planning, Scheduling and Coordination, Renewable Energy Sources, Plug-in Electric Vehicles
Abstract: The rapid proliferation of electric vehicles (EVs), empowered by vehicle-to-grid (V2G) technology, provides a novel and scalable form of flexible energy storage. Integrating V2G-enabled EVs into community microgrid (C-MG) presents a promising solution to mitigate supply-demand imbalances induced by renewable energy sources. However, the optimal scheduling of C-MG with integrated V2G energy storage remains significant challenges, primarily due to modeling complexities arising from large system scale and stochastic user behaviors, as well as the non-anticipativity and all-scenario-feasibility requirements for optimal decision-making. To address these challenges, this paper proposes a multi-stage robust optimal method to obtain the non-anticipative all-scenario-feasible scheduling decisions for C-MG with V2G energy storage. Specifically, an EV aggregation model that incorporates uncertainties in charging energy demand and large-scale heterogeneous charging behaviors is given first. Then an exact aggregated robust feasible region (ARFR) for scheduling decisions is developed, which guarantees non-anticipativity and all-scenario-feasibility. Based on this ARFR, the optimal charging/discharging profiles for the charging station are determined by solving an equivalent mixed integer linear programming (MILP) problem aimed at minimizing operational costs. Comparative case studies across different methods, EV types and scaling cases confirm that the proposed method enhances stable microgrid operation under various uncertainties.

Keywords: Planning, Scheduling and Coordination, Robust Manufacturing, Machine learning
Abstract: Mixed hot-and-cold charging improves energy efficiency in hot rolling, but uncertain slab rolling times can make deterministic schedules fragile and lead to hot-charging deadline violations. This work-in-progress paper presents a learning assisted robust scheduling framework. A decoupled dual-output deep neural network predicts both nominal rolling time and an instance-level uncertainty half-width. The predicted intervals are embedded into a budgeted polyhedral uncertainty set, and robust time recursion is used to evaluate schedules under worst-case rolling-time delays. An adaptive large neighborhood search heuristic then constructs feasible schedules within practical computation time. Preliminary experiments on industrial rolling data show that the decoupled architecture improves the balance between point accuracy, interval coverage, and interval sharpness. Monte Carlo validation further indicates that a small robustness budget can remove most deadline violations with limited cost increase.

Keywords: Planning, Scheduling and Coordination, Semiconductor Manufacturing, Optimization and Optimal Control
Abstract: We address a semiconductor cluster tool scheduling problem under just-in-time (JIT) time-lag constraints. Wafers follow route-dependent and sometimes reentrant flows through alignment, processing, chamber cleaning, and final cooling under blocking, no-overtake, valve mutual exclusion, and bounded time-lag constraints. We formulate the problem as an extended flexible blocking job shop with state-dependent cleaning and propose a clean-aware Constraint-Programming (CP)-guided large neighborhood search (LNS) framework. The method combines a constructive no-wait initial schedule with destroy-and-repair search, where CP jointly optimizes machine assignment, local resequencing, blocking occupancy, and mandatory cleaning decisions. Family-switch cleaning and count-triggered cleaning are handled inside CP repair, while idle-triggered cleaning is applied after realized gaps become known. The data-guided destroy-repair search uses online search-performance signals to steer neighborhood selection. An event-based decoder and validator convert each accepted machine-level schedule into executable MoveList actions and enforce detailed equipment-level rules. Experiments on eight benchmark instances show that the method generates executable schedules in all tested instances and outperforms existing methods on six of the eight instances.

Keywords: Planning, Scheduling and Coordination, Agricultural Automation, Task Planning
Abstract: Straw, as a byproduct of agricultural production, is used in resource-oriented, feed-oriented, and substrate-oriented ways, while straw recycling is the first step toward its efficient utilization. In practice, straw recycling enterprises first determine the baling plan and then the transport plan. Such a sequential decision-making method increases the time delay between these two processes and the maximum completion time. This study proposes an integrated scheduling framework to synchronize straw baling and transportation by jointly determining the working routes of straw balers and transporters, and proposes an efficient solution algorithm to solve it. Numerical experiments using real-world datasets show that our proposed methodology provides high-quality solutions within a reasonable computation time and reduces the maximum completion time, compared with the sequential decision-making method. Expanding the fleets of straw balers and transporters shortens the maximum completion time, and more straw balers also significantly decrease waiting time.

Keywords: Agricultural Automation, Foundations of Automation, Sensor Networks
Abstract: With the advancement of agricultural mechanization and the increasing size of field machinery, tire-induced soil compaction has become a major constraint on soil structure, water transport, and crop production. Tire inflation pressure is a key tire parameter because it changes tire–soil contact conditions and acts jointly with wheel load and travel speed. In this study, a soil-bin experiment was conducted using a 12.4–48 bias-ply tire to investigate the effects of tire inflation pressure on soil physical properties, hydraulic properties, aggregate stability, and pore structure under different load and speed conditions. A total of 45 treatments were established, including five tire inflation pressures, three loads, and three travel speeds. Soil samples were collected using a 3 × 3 spatial matrix consisting of three lateral wheel-track positions and three soil layers, namely 0–10, 10–20, and 20–30 cm. Bulk density, soil penetration resistance, total porosity, saturated hydraulic conductivity, gravimetric water content, field capacity, maximum water-holding capacity, and aggregate stability indices were measured. Nuclear magnetic resonance and X-ray computed tomography were further used to characterize soil pore structure in representative samples. The results showed that increasing tire inflation pressure and load, together with decreasing travel speed, intensified soil compaction. The strongest responses to tire inflation pressure occurred mainly within 150–250 kPa, whereas further changes became weaker at 300–350 kPa. Across the wheel-track width, the middle position showed greater compaction than the two side positions, and this lateral difference was most pronounced at 200–250 kPa. Vertically, the 0–10 cm soil layer was consistently more affected than the 10–20 and 20

Keywords: Agricultural Automation, Human-Centered Automation, Human Factors and Human-in-the-Loop
Abstract: Agricultural production imposes strict seasonal demands with tight production schedules. Concurrently, as the level of agricultural mechanization continues to increase, agricultural machinery operators are required to perform long periods of continuous work under high intensity. Their operational state significantly impacts the efficiency and effectiveness of mechanized operations. Existing research focuses on driver fatigue in road transportation, overlooking operator fatigue in agricultural machinery contexts. This study aims to investigate the environmental antecedents of fatigue in agricultural mechanization. Specifically, it explores the correlation between operator fatigue and water-soil environmental factors based on evidence from seeding. The research selected 800 farmland soil samples and measured soil compaction and moisture content at the surface, middle, and deep layers. Additionally, EAR and MAR values of the machinery operators were captured. These metrics were used to derive blink frequency and yawn frequency. Operator fatigue was detected using the PERCLOS algorithm. Pearson correlation analysis was employed to explore the correlation between operator fatigue levels and the water-soil environment. The findings provide a theoretical basis for formulating operator assignment plans and operational safety warning strategies.

Keywords: Manufacturing, Maintenance and Supply Chains, Intelligent and Flexible Manufacturing, Cyber-physical Production Systems and Industry 4.0
Abstract: To address the challenges faced by small and medium-sized agricultural machinery manufacturers, including dispersed manufacturing capabilities, weak cross-enterprise collaboration, and inadequate responsiveness to orders, this study proposes a social manufacturing–oriented organizational paradigm tailored to the agricultural machinery industry. First, the concept and operational characteristics of agricultural machinery social manufacturing are clarified, and a system framework is established that covers manufacturing resource digital modeling, order-driven community formation, and collaborative execution and delivery. The key operating mechanisms are then elaborated from the perspectives of information flow, capital flow, and manufacturing-process collaboration. Building on this foundation, a multi-task-domain evaluation and community screening method for manufacturing resources is developed to support resource combination and community selection under order-driven conditions. Finally, an orchard/hilly-area narrow-track multifunctional tractor is used as a representative scenario to analyze the community formation and screening processes for component supply, assembly, inspection, and logistics resources, thereby verifying the feasibility and engineering applicability of the proposed approach.

Keywords: Diagnosis and Prognostics, AI-Based Methods, Agricultural Automation
Abstract: As core intelligent equipment in agricultural harvesting, safe and efficient operation of rice combine harvesters is crucial for stabilizing yield, ensuring supply and reducing maintenance costs. Pre cise predictive fault diagnosis technology is vital to boost their efficiency, safety and service life. While l arge language models (LLMs) show strong capabilities in fault diagnosis, they suffer from inherent dra wbacks: unreliable generation results, slow response, high deployment costs and poor scalability. To add ress these issues, this paper proposes a fault diagnosis method for rice combine harvesters integrating a knowledge graph (KG) and mixture-of-experts (MoE) model. First, a KG is built to determine the opti mal fault location–cause–solution path, improving diagnostic accuracy. Second, a multi-dimensional retrie val method combining vector retrieval, BM25 retrieval and graph retrieval is introduced to enhance con text awareness and fault feature matching accuracy. Finally, a sparsely activated MoE mechanism is des igned, comprising a single-signal expert subnet (for single fault feature identification) and a fused-signal expert subnet (for multi-source signal collaborative analysis). This reduces resource consumption while enhancing scalability and real-time response. Comprehensive tests are conducted on diagnostic accuracy, inference latency, response efficiency, model scale, computational complexity and generalization. Results s how the proposed method significantly outperforms traditional and single LLM-based methods, with imp roved accuracy, reduced time/resource consumption and enhanced multi-scenario adaptability, providing reliable technical support for efficient rice combine harvester fault diagnosis.

Keywords: Human-Centered Automation
Abstract: Simulation-based training helps control room operators maintain decision-making and intervention abilities in highly automated process plants. Operator training simulators provide a virtual representation of plant processes and control systems to execute training scenarios, including abnormal operating conditions. Effective training requires diverse scenarios aligned with operator experience and defined learning objectives. In practice, creating scenarios for a specific plant is labor-intensive and requires specialized expertise, especially after plant or control system changes. This motivates automating scenario generation based on plant operating procedures, possible fault conditions (e.g., pump failure during start-up), the operator's training history, and didactic concepts such as learning objectives and competencies. We present a model-based assistance system that integrates control logic, fault conditions, the operator's training history, and formalized didactic concepts in a knowledge graph based on an ontology. The approach includes competency-oriented prioritization based on a simplified skill-decay heuristic to identify training needs and guide scenario generation. A feasibility example demonstrates the generation of admissible training scenarios from the integrated knowledge graph using prioritization and ontology-defined constraints.

Keywords: Agent-Based Systems, Intelligent and Flexible Manufacturing, Robust Manufacturing
Abstract: Modern production systems face increasing demands for reconfigurability, driven by product variability and shortened innovation cycles. In multi-agent system (MAS) architectures, the capability assumptions that underlie certification decisions may become invalid during operation due to reconfiguration or parameter changes. Traditional verification, however, relies on static, design-time models. This paper proposes runtime dependability certificates (RDCs), which are machine-interpretable contracts whose validity is continuously enforced at runtime and automatically revoked when certified conditions are violated. The framework separates probabilistic, large language model (LLM)-based capability extraction from a deterministic certification core. This ensures that automated knowledge extraction does not compromise formal dependability guarantees. Certificate validity is enforced through three identity gates that cover process input binding, resource identity attestation, and continuous runtime monitoring. A laboratory case study combining additive manufacturing and robotic handling demonstrates that RDCs enable agents to detect capability violations while avoiding unnecessary reassessment and maintaining coordination efficiency.

Keywords: Probability and Statistical Methods, Human Factors and Human-in-the-Loop, Cyber-physical Production Systems and Industry 4.0
Abstract: Human–machine interaction is a well-researched topic and is highly relevant to modern production systems and the concept of Industry 5.0. Technical systems, including their interactions with humans, often use digital twins (DTs) to better understand and manage technical processes. Although the technical structure of such systems can typically be modeled with relatively high accuracy using DTs, since many machine parameters are known, humans are subject to considerable uncertainty. This uncertainty becomes even more pronounced when several DTs are aggregated into a hierarchical DT framework.

This paper presents a hierarchy-aware framework for uncertainty quantification (UQ) and uncertainty propagation (UP) in cascaded DTs. Uncertainty is modeled at the component level as aleatoric and epistemic, including assumption-consistency checks to detect domain-of-validity violations. Sensitivity-based screening identifies variance-dominant components, enabling contribution-driven allocation of UQ fidelity. The framework is demonstrated using a representative human–robot collaboration (HRC) case study with three coupled DTs. The results indicate that uncertainty dominance is mainly determined by structural coupling rather than by local model accuracy. The findings establish uncertainty dominance as a structural property of hierarchical DTs and provide a method for scalable, reliable uncertainty management in human-centered cyber-physical systems.

Keywords: Computer Vision in Automation, Sensor-based Control, Sensor Networks
Abstract: In fulfillment centers, diverse objects move continuously from inbound to outbound operations and can become jammed due to excessive conveyor friction, incorrect orientation, or mechanical failures. Traditional jam detection approaches rely on object detection models to identify objects, followed by tracking algorithms (such as IoU overlap and Kalman filtering) to monitor motion over time. This pipeline requires thousands of manual annotations, consuming approximately two weeks of effort, and is limited to annotated object classes. We present a training-free, object-agnostic jam detection method that eliminates the need for labeled data. Our approach uniformly samples reference points within the monitoring region when no objects are present. As objects occlude these points, we detect motion. When a sufficient fraction remains occluded beyond a temporal threshold, we classify the event as a jam. Unlike conventional point tracking—which treats occlusion as a failure case—our approach repurposes occlusion as a detection signal, monitoring whether reference points remain persistently occluded rather than tracking where they move. Our experimental evaluation on 1,069 videos demonstrates that AllTracker achieves 100.00% precision and 93.33% F1 score, significantly outperforming classical sparse tracking methods while maintaining training-free deployment. This approach offers three key advantages: (1) no training data or manual annotations, (2) object-agnostic generalization to arbitrary object types, and (3) significantly reduced development time.

Keywords: Computer Vision in Automation, Deep Learning in Robotics and Automation, Simulation and Animation
Abstract: This paper presents CADetect, a fully automated pipeline for generating synthetic object detection datasets directly from CAD models. The workflow imports and normalizes STL geometry, generates scenes, samples camera viewpoints on a sphere, applies domain randomization, renders images, and creates YOLO-compatible bounding box annotations automatically. To investigate strategies for reducing the reality gap between synthetic renderings and real imagery, the dataset generation was developed iteratively by gradually increasing synthetic variance, including viewpoint coverage, appearance variation, camera parameter perturbations, compression effects, and background and lighting diversity. To evaluate the suitability of the resulting synthetic data, several model variants from two recent generations of YOLO models under identical hyperparameter conditions were trained and compared. The experiments show that all evaluated models converge to a similar detection accuracy on the presented dataset, while computational costs differ substantially across model variants. Based on the observed trade-off, smaller model variants provide the most efficient inference feasibility for resource-constrained environments. The results indicate that a reproducible, scalable synthetic data pipeline can serve as a viable foundation for training object detection models and for systematically studying the impact of data variability on model robustness.

Keywords: Intelligent and Flexible Manufacturing, Machine learning, Optimization and Optimal Control
Abstract: This paper proposes Vibration Control Optimization, a method that automatically optimizes vibration configurations on flexible feeders. The flexible feeder features a vibration plate that reorients parts via configurable vibrations. Compared to vibratory bowl feeders, the flexible feeder has favorable properties for high-mix, low-volume production due to its significantly faster changeover time. However, instantaneous changeover assumes that pre-tuned vibration configurations are available. Vibration Control Optimization offers real-world, fast, automatic vibration configuration optimization by leveraging the pose estimator Spyropose and employing Bayesian Optimization with Wilson-Score Kernel Density Estimation as a function estimator. The method is implemented on an Asycube 530 flexible feeder. The results demonstrate the method's effectiveness, with convergence within 2 hours of real-world sampling across all tested parts. Analysis of results shows a significant increase in the likelihood of the target orientation for all converged configurations.

Keywords: Intelligent Transportation Systems, Sustainability and Green Automation
Abstract: Estimating traffic carbon emissions from video on edge devices is challenging due to the heavy computing load required for both visual perception and dynamic modeling. This paper presents a hybrid estimation framework designed to address this computational bottleneck. In the perception stage, we use a Background-Mask Vision Transformer (BM-ViT) to reduce redundant calculation. By applying a temporal background model to mask stable areas of the road, the system focuses processing resources only on moving vehicles.In the modeling stage, the framework groups similar vehicle trajectories into representative clusters. Instead of running physical simulations for every individual vehicle, we apply a Vehicle Specific Power (VSP) model combined with Monte Carlo simulations only to these cluster centers. A fallback mechanism is also included to monitor clustering quality, ensuring the system calculates emissions individually when traffic behavior is too complex to group. Experimental results indicate that this approach reduces the visual computational resources by 65.2% and decreases emission-modeling computational load by approximately 81.7%,while keeping the global relative error below 1.2% compared to complex physical simulations.

Keywords: Inventory Management, Demand Side Management, Machine learning
Abstract: Accurately predicting subsequent demand dynamics from historical sales and pricing actions is critical for effective replenishment and pricing decisions in e-commerce systems. However, they contain excessive zero-demand observations, which can distort the overall demand distribution. Recently, recurrent neural networks (RNNs) have proven effective in capturing temporal demand dynamics, but they do not guarantee economically consistent price-demand relationships that are crucial for pricing decisions. To overcome this limitation, we propose Monotonic Integrated Recurrent Representation Of probabilistic Risk (MIRROR), a price-constrained probabilistic demand forecasting framework to support joint replenishment and pricing decisions. MIRROR combines autoregressive temporal demand modeling with monotonic neural components that enforce partial monotonicity with respect to price, while modeling demand uncertainty through a zero-inflated probabilistic formulation. Empirical results on a real-world benchmark demonstrate that MIRROR guarantees economically consistent price-demand monotonicity while achieving improved probabilistic forecasting performance over unconstrained RNNs.

Keywords: Learning and Adaptive Systems, Machine learning, AI-Based Methods
Abstract: 部署边缘多对象跟踪（MOT）系统 计算设备面临着严重的冲突 受限的计算资源与实时 要求。而现有的基于检测的跟踪 范式提供高精度，但常常存在以下问题 处理高帧率视频时显著的延迟 由于探测器开销过高。大多数 当前边缘MOT系统依赖静态均匀采样， 这忽略了视频的时空动态 内容，导致精度或资源的下降 浪费。本文提出了CAKS-MOT，即实时技术 基于内容感知的边缘MOT加速框架 自适应关键帧采样。该框架引入了 创新的重要性预测方法与智能传感 自适应采样技术以平衡关键帧 内容覆盖的重要性。广泛的实验 在多个基准测

Keywords: Machine learning
Abstract: Polyp image segmentation plays an essential role in the prevention of colorectal cancer. However, due to the complex morphology of polyps and their low contrast against the background, existing methods often face challenges related to missed detection and false detection. In order to surmount these issues, this study proposes a polyp segmentation algorithm that integrates spatial-domain and frequency-domain information, employing an error-feedback-guided mechanism to facilitate interactive optimization between two domains. Specifically, spatial-domain processing part utilizes an adaptive residual ConvNeXt module to fully capture local details and spatial contextual information from image. Frequency-domain processing part introduces discrete wavelet transform and Chebyshev polynomial-based KAN convolution to perform frequency decomposition and modeling, explicitly leveraging high-frequency edge features and low-frequency structural information. To further organize features from both domains, an error-feedback-guided information interaction module is designed to iteratively compute the information loss between the fused result and the original features, thereby dynamically generating optimized fusion weights. Furthermore, ReduceLROnPlateau strategy, which enhances training stability and mitigates overfitting, is adopted for adaptive learning rate adjustment. Experimental results demonstrate that proposed method effectively detects polyp tissue, validating the performance of synergistic spatial-frequency domain modeling in medical image segmentation.

Keywords: Task Planning, Planning, Scheduling and Coordination, Cloud Computing For Automation
Abstract: This paper studies the virtual machine placement problem under cluster power budget constraints in cloud data centers. In this problem, VMs must be assigned to physical servers subject to server-level computing and storage capacity limits, while the total power consumption of servers within each cluster is restricted by a cluster-level power budget. These cluster-level power budgets introduce additional coupling among server placement decisions, which significantly increases the complexity of the problem. The objective is to minimize the total cost, consisting of the penalty cost of unassigned VMs and the heterogeneous fixed activation cost of servers, thereby capturing both service loss and server heterogeneity. To solve the problem efficiently, we first formulate it as an integer programming model. Based on the problem structure, we then develop a Dantzig–Wolfe decomposition framework. The resulting pricing subproblem is formulated as a twodimensional knapsack problem, which is solved by a labeling algorithm. To further accelerate the solution process, we develop a column-generation-based matheuristic, which combines a linear-relaxation-solution-guided construction heuristic for generating high-quality initial columns and a diving-based fixing procedure for obtaining high-quality integer feasible solutions. Computational experiments on synthetic instances demonstrate the effectiveness and scalability of the proposed approach.

Keywords: Task Planning, Reinforcement, Cloud Computing For Automation
Abstract: To address the multi-objective optimization challenge in balancing local and edge computing rate within Wireless Powered Mobile Edge Computing (WPMEC), this paper proposes a Neighbor-Assisted Multi-Agent Deep Reinforcement Learning framework for multi-objective computation offloading (NAMADRL). The framework employs a decomposition strategy to transform the original problem into multiple subproblems with specific preference weights. By introducing a neighbor-assisted collaboration mechanism, agents with similar preferences can share action information and optimize scalarized rewards, thereby effectively accelerating policy convergence. Experiments demonstrate that compared to the traditional MOEA/D algorithm, NAMA-DRL exhibits significant advantages in terms of convergence speed, solution set distribution uniformity, and robustness.

Keywords: Telerobotics and Teleoperation, Manipulation Planning, Force and Tactile Sensing
Abstract: In this article, we propose system (name omitted for submission), a novel joint-matching teleoperation system designed to enhance dataset collection for manual and bimanual tasks. System is specifically tailored for controlling the manipulators and features a custom controller with force feedback and adjustable sensitivity modes, enabling precise and intuitive operation. In addition, it integrates a user-friendly dataset recording interface, simplifying the process of collecting high-quality training data for imitation learning. System is designed to be reliable, costeffective, and easily reproducible, making it an accessible tool for researchers, laboratories, and startups engaged in advancing robotics through imitation learning. The architecture of the system, based on specially developed units, is reconfigurable and can be adapted to different types of manipulator kinematics. Experimental results demonstrate that system can be effectively utilized for large-scale data collection to train VLA models, while user studies highlight the importance of adaptive sensitivity adjustment and force feedback for efficient, high-quality data acquisition. Assembly instructions, hardware descriptions, and source code will be provided at https://github.com/hiddensubmission/Teleoperation-System.

Keywords: Machine learning, Robust/Adaptive Control, Data fusion
Abstract: Twin Support Vector Machine (TWSVM) works well on high dimensional, small sample classification tasks. But traditional TWSVM is very sensitive to noise and outliers. This can cause unstable decision boundaries when the data is complex. To fix this problem, we propose a new model called SoftCapped L1 Twin Support Vector Machine (WSCL1-TWSVM). This model uses a Welsch Soft-Capped L1 Loss. It makes the model stronger against noise. It also improves the stability of decision boundary learning and works better with high dimensional data.We tested the model on ten benchmark datasets and on LUAD RNASeq data. WSCL1-TWSVM consistently outperforms the classic TWSVM and its robust variants. It gives better accuracy, stronger noise resistance, and more stable results.When noise is added, the model's performance drops slowly and in a more controlled way. The drop rate is only 0.0928, which is much lower than other methods.On the benchmark datasets, the average accuracy is 0.8589 on clean data and 0.8038 under cross noise conditions. The model ranked Top-1 in 18 out of 50 settings and Top-2 in 35 settings.In the LUAD cohort screening, the model achieved ACC of 0.9517, F1 score of 0.9541, and AUC of 0.9859. The high confidence overlap rate reached 0.5368. This shows better consistency and reliability in patient stratification. WSCL1-TWSVM can serve as a robust frontend tool for precision medicine. It provides stable and reproducible samples for downstream analysis.

Keywords: Planning, Scheduling and Coordination, Motion and Path Planning, Automation in Construction
Abstract: This paper studies online multi-robot exploration and line coverage in unknown environments, where robots must explore the environment and service features without prior topology knowledge. The main challenge is that the feature graph is revealed incrementally, causing routing structure and workload allocation to update during execution. To address this, we propose an iterative hierarchical cyclic merging regulation (iHCMR). iHCMR is a coupled exploration-coverage framework that constructs the feature graph from local sensing, generates local exploration and servicing trajectories, and replans coverage tasks as new features are discovered. iHCMR preserves Morse-bounded tours throughout graph discovery, enabling newly detected features to be integrated without introducing coverage discontinuities. We prove coverage completeness and bounded competitiveness relative to the offline optimum. Simulations show improved workload balance and reduced makespan, compared with the graph-based and greedy online strategies.

Keywords: Collaborative Robots in Manufacturing, Optimization and Optimal Control, Industrial and Service Robotics
Abstract: This paper presents a Reflexive Path-Following MPC for multi-arm cooperative manipulation using a Hybrid Collision-Aware Modulation strategy. By embedding a path variable into a large-scale optimal control problem, we achieve spatial-temporal decoupling. Collision safety is efficiently evaluated via Bounding Volumes for the payload and Signed Distance Fields for the articulated arms. Despite this complex optimization landscape, the implementation supports a 10 Hz planning update rate in hardware experiments, with observed solver latencies compatible with the 100 ms control interval. Hardware experiments validate our approach through an ablation study across three distinct dynamic interference scenarios, demonstrating robust, redundancy-exploiting whole-body avoidance.

Keywords: Deep Learning in Robotics and Automation, Model Learning for Control, Autonomous Agents
Abstract: Vision-language-action (VLA) models are a promising foundation for general-purpose robot manipulation, but their inference latency remains a major obstacle to stable real-time control. Asynchronous execution alleviates this bottleneck by overlapping inference with action execution, yet it also creates an observation-action mismatch: the next action chunk is predicted from a stale observation while the robot continues to move. Directly stitching consecutive chunks therefore introduces handoff discontinuities, action jitter, and downstream failures in contact-rich manipulation. Existing remedies typically require either full-policy retraining or architecture-specific runtime logic. We propose Action ControlNet (ACNet), a lightweight delay-aware adapter that conditions a largely frozen action head on the motion suffix currently being executed. ACNet injects residual condition features into the action expert, preserving the pretrained backbone while remaining compatible with generative action heads such as diffusion-based and flow-matching models. Experiments on Kinetix and Meta-World MT50 show that ACNet consistently improves robustness under delay and reduces trajectory jerk with modest overhead. On Kinetix, ACNet improves success at delay d=4 by 22 points over naiveasync while using only sim20% of the trainable parameters required by Training-RTC. On Meta-World, ACNet improves success at every tested delay and yields substantially smoother asynchronous transitions on a single RTX 4080 SUPER.

Keywords: Deep Learning in Robotics and Automation, Reinforcement, Model Learning for Control
Abstract: Autonomous lunar geological sampling requires both open-ended semantic reasoning and high-precision physical control. While Vision-Language-Action (VLA) models excel at semantic planning, their susceptibility to extreme High-Dynamic-Range (HDR) lunar illumination and low execution frequency often lead to fatal misalignments and "dynamic evasion" during microgravity contact. To address this, we propose VLA-ResGrasp, a decoupled hierarchical control framework. The system explicitly separates a macro-level VLA planner -- fine-tuned via QLoRA for global semantic approach trajectories -- from a micro-level Residual Reinforcement Learning (RL) controller. Operating at high frequency, the RL module employs an asymmetric DDPG agent driven by localized spatiotemporal grayscale features to perform closed-loop geometric corrections, effectively mitigating both HDR visual artifacts and dynamic uncertainties. Extensive procedural evaluations demonstrate that VLA-ResGrasp achieves an 85.0% success rate in robust grasping tasks, significantly outperforming monolithic VLA baselines, and maintains a highly reliable 70.0% success rate in complex long-horizon obstacle-avoiding collection in unstructured extraterrestrial environments.

Keywords: Health Care Management, Petri Nets for Automation Control
Abstract: Data-driven clinical pathway models are increasingly used to support performance analysis and operational decision making, but their dependence on finite event logs may introduce uncertainty in downstream performance estimates. This paper addresses the problem of assessing the variability of throughput predictions in Probabilistic Timed Petri Nets (PTPNs) identified from clinical pathway data. We propose a practical procedure to quantify how uncertainty in mined routing probabilities propagates to throughput estimates. Starting from a PTPN, we generate hypothetical scenarios by sampling branching probabilities from Dirichlet distributions and evaluate their impact on throughput using standard performance-analysis pipelines. The resulting dispersion is summarized through a relative stability index and analyzed with respect to the stability threshold, initial workload, and global scaling of transition durations. The approach is illustrated on an anterior cruciate ligament (ACL) clinical pathway. The results show that throughput uncertainty decreases as the assumed routing support increases, whereas the absolute throughput level depends on workload and timing assumptions. The analysis provides a reliability check for interpreting throughput predictions obtained from finite clinical event logs.

Keywords: Hybrid Strategy of Intelligent Manufacturing, Process Control, Machine learning
Abstract: The tool-workpiece engagement dynamics in milling involve abrupt interaction forces that can degrade manufacturing performance. This paper presents a machine learning-based real-time digital twin, iTwin, for end-mill deformation estimation and chatter suppression. By integrating cutting force sensing, milling dynamics, and a learning-based surrogate model, the proposed framework estimates the real-time end-mill deformation response for process inspection and adaptive control. Offline validation at representative-point and whole-dynamics levels shows agreement with laser-measured maximum displacement and finite element analysis (FEA)-predicted nodal displacement fields. Online experiments on different aluminum materials and toolpath geometries verify that the iTwin captures the transition from chatter to stable cutting after control and consistently improves surface roughness. These results demonstrate the effectiveness of the proposed iTwin for real-time tool-state estimation and chatter-aware machining-quality improvement.

Keywords: Actuation and Joint Mechanisms, Compliant Joints and Mechanisms, Robust/Adaptive Control
Abstract: This paper presents the Hart-based Parallel-pinching Adaptive gripper, termed the HPA gripper, a compact underactuated gripper intended for robotic grasping tasks that require both controlled fingertip motion and passive adaptation to object geometry. Existing underactuated grippers provide simple and robust grasping solutions, but it remains challenging to integrate parallel pinching and adaptive enveloping in a single compact mechanism without additional actuators, sensors, or complex control logic. The proposed design addresses this challenge by coupling a modified Hart mechanism with a parallelogram constraint and a lost-motion transmission. During the approach stage, the modified Hart mechanism guides the distal endpoint along a near-horizontal path, while the parallelogram constraint maintains a suitable fingertip posture for parallel contact. After object contact, the lost-motion transmission redirects the remaining input motion and triggers inward rotation of the distal phalanx, enabling adaptive enveloping. Kinematic analysis was conducted to locate the exact straight-line point and evaluate the straightness of the practical fingertip endpoint. A prototype was fabricated to verify the mechanical feasibility of the concept through representative demonstrations, including small-range parallel pinching, wide-range parallel pinching, and adaptive enveloping. The demonstrations indicate that the HPA gripper can realize dual-mode grasping through passive mechanical coordination, offering a simple and compact solution for low-cost robotic manipulation tasks with limited sensing and control resources.

Keywords: Agent-Based Systems, Intelligent and Flexible Manufacturing, Industrial and Service Robotics
Abstract: Automation for robotic system configuration remains an open challenge in Industry 5.0. The inherent uncertainty of incomplete requirements, dynamic process flows, and execution-level constraints fundamentally exceed the assumptions of conventional approaches. To address these limitations, this paper proposes a skill-augmented large language model (LLM) agent for simulation-based robotic system layout design. The agent interprets incomplete user requirements by generating layout-determining sub-questions and resolving them into structured layout specifications. These specifications are subsequently transformed into executable 3D environments that are programmatically instantiated within an open-source simulator. A hierarchical domain-specific skill library, governed by a progressive disclosure mechanism, enables controlled injection of manufacturing expertise and allows the agent to dynamically invoke callable skills throughout the design process. The generated layouts are automatically validated through robotic simulation, thus forming an end-to-end pipeline. A case study on 3D bin-packing workcell design demonstrates the feasibility and practicability of the proposed approach. The results highlight the potential of skill-augmented LLM agents towards enabling largely autonomous and execution-aware design of reconfigurable robotic systems.

Keywords: Agent-Based Systems, Motion and Path Planning
Abstract: Current Vision-Language Navigation (VLN) systems, lacking an effective representation of the spatial environment, suffer from inefficient and blind exploration in object navigation tasks. Existing statistical or mapping-based representations of spatial relationships often fail to fully align with human thinking. Therefore, this paper proposes a dynamic prediction zero-shot object navigation framework that represents the spatial environment in a hierarchical manner: "room-functional area-object." First, by constructing and updating a hierarchical spatial relationship graph, we achieve a spatial distribution model of the indoor environment that aligns with human thinking, enabling the agent to possess persistent spatial perception and memory capabilities. Next, based on the hierarchical spatial description, we parse language commands into a unified structured form for target location prediction. We propose a dynamic prediction mechanism based on Chain-of-Thought (CoT). Furthermore, we introduce an adaptive exploration strategy that supports adaptive search behavior in different scenarios. By leveraging deterministic search, fuzzy search, and a blacklist mechanism, we effectively reduce ineffective search behavior and improve search efficiency. Experimental evaluation results on public datasets such as HM3D and MP3D demonstrate that our approach achieves a 4.7% improvement in success rate (SR) over the best baseline method, while also achieving navigation behavior that better aligns with human behavior, demonstrating strong practicality and robustness.

Keywords: AI-Based Methods, Model Learning for Control, Deep Learning in Robotics and Automation
Abstract: Unmanned aerial vehicles (UAVs) equipped with manipulators enable aerial interaction tasks such as object pickup and transport in environments that are difficult for ground robots to access. However, reliable autonomous aerial grasping remains challenging due to limited onboard perception, dynamic flight conditions, and the need for precise visual alignment. This paper investigates the use of vision language models (VLMs) to guide UAV-based aerial manipulation through visual reasoning and command generation. We present a system that integrates a downward-facing camera, an onboard flight controller, and an external GPU-based VLM inference module to generate motion commands for object centered grasping. Two control strategies are explored: (1) direct prediction of low-level motion commands and (2) high level language-mediated control for iterative target alignment. A dataset of aerial manipulation samples is collected in simulation to fine-tune several open-source VLMs using LoRA- based parameter adaptation. The approach is evaluated in both simulation and real-world experiments using multiple object types and environmental conditions. Experimental results demonstrate that the fine-tuned models achieves the highest grasp success rate and improved centering accuracy compared with other evaluated models, highlighting the potential of VLM-guided perception and control for flexible aerial manipulation with UAV platforms.

Keywords: AI-Based Methods, Telerobotics and Teleoperation, Human Factors and Human-in-the-Loop
Abstract: Industrial automation is at a pivotal moment, as Physical AI is driving a transition from rigid, hand-engineered automation systems toward more flexible and adaptive systems. This shift has created a growing demand for large-scale, real-world robot demonstration data, making teleoperation an increasingly important mechanism for data collection. However, high-quality teleoperated demonstrations remain difficult to obtain in practice, as novice operators often produce episodes that are task-successful but suboptimal for downstream use due to inefficient motion, repeated corrections, gripper chatter, or operation near robot joint limits. We present a Data Quality Assessment and Feedback (DQAF) framework that closes the loop in teleoperation by providing immediate post-episode feedback grounded in semantic task progress and robot telemetry.

The framework extracts quality relevant signals such as sub-task progress, motion smoothness, stalls, kinematic limits and converts them into structured quality assessments and actionable natural-language feedback. Unlike binary success/failure feedback, the proposed system explains why an episode is suboptimal and highlights specific behaviors to correct in the next trial. We evaluate the framework through a diagnostic validation study and a pilot user study. In the validation study, the system is compared with a human reviewer during dataset curation, producing rejection reasons and actionable feedback for improvement. In the pilot study with three novice operators across two manipulation tasks, the operator who received the system’s immediate, automated post-episode feedback improved faster than those who did not, producing higher-quality demonstrations sooner.

Keywords: AI and Machine Learning in Healthcare, AI-Based Methods, Machine learning
Abstract: Alzheimer’s disease (AD) classification using magnetic resonance imaging (MRI) is essential for early clinical intervention. However, deep learning approaches often face challenges such as limited medical data, high intra-class variability, and uneven diagnostic contributions from individual MRI slices. To address these issues, we propose a hybrid framework that combines attention-guided multi-slice deep representation learning with a variance-weighted dictionary learning (VW-DL) classifier. A convolutional neural network first extracts slice-level features, which are aggregated through an attention mechanism to generate a volume-level representation. Class-specific dictionaries are then constructed in the deep feature space, and a variance-based weighting strategy is introduced to suppress unstable feature dimensions during sparse reconstruction. Experiments on the ADNI dataset with strict subject-level data partitioning demonstrate that the proposed method achieves an accuracy of 96.2%, outperforming the baseline attention-guided multiple instance learning (MIL) model. It also maintains a balanced sensitivity 95.2% and specificity 97.8%, effectively reducing false negatives and showing strong potential for reliable automated AD diagnosis.

Keywords: AI and Machine Learning in Healthcare, Deep Learning in Robotics and Automation
Abstract: Precision medical robotics demands adaptive decision-making under strict safety, interpretability, and execution constraints. Although recent Vision-Language-Action (VLA) models show strong multimodal reasoning ability, their continuous action generation paradigm is not well suited for precision medical tasks, where reliable closed-loop operation may also depend on non-action system function calls. To address this gap, we propose MedVLA, a hierarchical framework that couples high-level multimodal reasoning with low-level function-constrained execution. We further introduce a scalable multi-agent pipeline to generate skill-oriented chain-of-thought(CoT) data for structured training. Built on different multimodal large-model backbones, MedVLA consistently improves performance after fine-tuning, demonstrating the effectiveness of the proposed framework across model variants. Under identical initial conditions, we perform 100 closed-loop flexible electrode implantation trials. The results show that MedVLA achieves a 95.0% task success rate, substantially outperforming representative VLA baselines, including OpenVLA (8%) and pi_0 (15%), in accuracy, stability, and safety. These results indicate that structured reasoning with constrained function-level execution is a practical route toward deployable precision medical robotics.

Keywords: AI-Based Methods, Autonomous Agents, Intelligent Transportation Systems
Abstract: Industrial digital twins can track process state and support simulation. However, critical workflow transitions are still often executed by fixed steps even when the workpiece reaches the next station in different conditions. This paper explores the problem in an internal digital twin of a composite-forming workflow for PA6/CF UD-tape laminate sheets. The original workflow follows a fixed sequence of heating, press-based forming, and downstream single-point incremental forming (SPIF). We keep this workflow and insert a pre-press decision layer that selects one of three admissible actions: additional heating, gripper reduction, or press entry. The controller operates on a compact pre-press state defined by stiffness, estimated sheet tension, sheet temperature, surface temperature, and gripper count. We compare an LLM-only policy with an LLM safety policy, which adds a safety-aware layer integrated with a runtime safety shield that blocks unsafe press transitions under process limits. Evaluation of multiple simulated scenarios shows that both policies reduced the safety violation rate and improved stiffness quality compared to the fixed baseline. The safety-aware layer further eliminates unsafe press attempts by converting them into explicit refusal. These results indicate that bounded LLM-based decision support can strengthen industrial digital twins when grounded in process-state and runtime safety checks.

Keywords: AI-Based Methods, Cyber-physical Production Systems and Industry 4.0, Autonomous Agents
Abstract: Designing control algorithms for cyber-physical systems (CPS) requires balancing performance objectives against hard safety constraints. This design process currently relies on expert manual tuning or evolutionary optimization methods that treat the problem as a black-box search, requiring a very high number of evaluations with no insight into why candidates fail. Large Language Models (LLMs) can generate and reason about algorithmic code, but they may lack awareness of physical safety boundaries, especially during runtime. We present MCP-DT (Model Context Protocol–Digital Twin), a domain-agnostic framework in which an LLM iteratively synthesizes control algorithms by conversing with a physics-based Digital Twin through the Model Context Protocol. During a candidate algorithm runs, if a constraint violation occurs, the Digital Twin returns enriched feedback, identifying which constraint was violated, at what system state, and what the safe operating boundary is, enabling the LLM to make targeted, reasoning-driven corrections. The engineer specifies only a goal, declarative safety constraints, and an algorithm interface. The paper validates MCP-DT on a battery fast-charging case study, demonstrating convergence into a feasible working solution.

Keywords: Automation in Life Science: Biotechnology, Pharmaceutical and Health Care, Modelling, Simulation and Optimization in Healthcare, Robotics and Automation in Life Sciences
Abstract: Suturing is a critical yet tedious aspect of nearly all surgical procedures. Proper suture placement is necessary to evenly distribute tissue forces, which promotes healing, minimizes complications, prevents leakage and tissue necrosis, and reduces scarring. We present Suture-It: a two-stage augmented dexterity framework for generating three-dimensional suture placement plans for nonplanar variable-width, non-branching wounds. In the first stage, a disparity map is used to reconstruct the wound and its surrounding tissue as a triangulated surface in three dimensions and to fit an adaptive-width ribbon to the segmented wound. In the second stage, the closure and shear forces along the ribbon are locally optimized using the Nelder-Mead algorithm, guided by a physics-based elliptical tissue force model. Physical experiments with chicken skin suggest that Suture-It produces admissible suture plans that can fully close wounds and avoid entry and exit point bunching, with minimal fluid leakage.

Keywords: Process Control, Machine learning, Robust Manufacturing
Abstract: Conventional process control and optimization methods have largely been developed for settings in which product quality can be monitored during the process or can be sufficiently characterized by static operating conditions. Under these assumptions, feedback regulation and constrained optimization provide effective tools for maintaining desired process behavior. However, these assumptions do not hold in several manufacturing processes where endpoint quality is available only after process completion and depends on the cumulative effect of a length-varying input profile. This paper studies the setting as a profile optimization problem under delayed quality feedback. We consider a framework that maps length-varying profiles into a low-dimensional B-spline parameter space, learns a surrogate model for the delayed quality response and performs profile optimization in the resulting parameter space. A synthetic dataset is used to examine the optimization behavior in a controlled setting, and a real-world food processing dataset is used to illustrate the same formulation in practice. The results show that the framework can reconstruct profiles that remain qualitatively similar to observed profiles while operating in a fixed-dimensional search space. These observations indicate that compact profile parameterization can serve as a practical tool for profile optimization problems with delayed quality feedback.

Keywords: Sensor-based Control, Compliant Assembly, Motion Control
Abstract: This paper presents a generic model-based methodology for defining and executing robust force‑based insertion skills capable of handling geometric uncertainties in robotic assembly tasks with sub-millimiter clearance fits. The approach can deal with geometric variability (i.e. different object shapes), insertions with multiple contact points such as dual pegs and complex mating features such as snap‑fit assemblies. Robustness is achieved through a sequence of task‑space degree‑of‑freedom reductions, one at a time, each executed via force‑based control and monitoring to systematically reduce uncertainties. In addition, the paper introduces an algorithm that automatically generates insertion skills for planar polygonal objects, which are commonly found in industrial applications such as printed circuit board (PCB) assembly and quality control. The methodology is validated through extensive experiments involving objects of varying geometries and tight sub‑millimeter clearance fits.

Keywords: Simulation and Animation, Machine learning, Process Control
Abstract: High-fidelity two-phase CFD for resin molding is too computationally expensive for iterative process design, despite the fact that filling quality and residual voids strongly govern yield and reliability. We propose an autoregressive neural PDE surrogate that predicts the spatiotemporal evolution of the resin volume-fraction field in 2-D cavity filling with non-Newtonian (Cross-power-law) rheology. The model uses a grid-based encode--process--decode architecture and is conditioned on rheological parameters, inlet volumetric flow rate, and gate geometry, thereby extending prior two-phase surrogate formulations largely limited to Newtonian settings. To better capture defect formation, we introduce an interface-aware loss that emphasizes the narrow resin–air transition region where void entrapment is determined. Using 1,000 OpenFOAM VOF simulations generated via Latin hypercube sampling, we benchmark four processor backbones (dilated ResNet, U-FNO, FNO, and U-Net) and find that dilated ResNet provides the most accurate long-horizon rollouts, while the interface-aware loss improves void-related accuracy without degrading bulk field predictions. The resulting surrogate delivers approximately four orders of magnitude speedup over CFD. Moreover, process KPIs derived from rollouts—filling time (R^2=0.99) and residual void ratio (R^2=0.75)---outperform a feature-based ExtraTrees baseline, particularly for void prediction (R^2=0.56). Overall, learning full spatiotemporal dynamics enables accurate, interpretable iterative resin molding process design—capabilities unattainable by CFD alone or scalar regression.

Keywords: Control Architectures and Programming, Smart Grids, Automation Technologies for Smart Cities
Abstract: Multi-agent reinforcement learning (MARL) faces challenges in capturing the spatial interdependencies and physical constraints of electrical networks. Existing paradigms often treat agents as isolated entities, resulting in fragmented perception and suboptimal coordination. To address these limitations, we propose the Graph-Enhanced Multi-Agent Transformer (GMAT) for real-time voltage stability control. Our methodology constructs a physics-informed virtual topology centered on critical buses, utilizing voltage sensitivity-based electrical distances to adaptively extract spatio-temporal voltage-power dynamics. These embeddings are processed by a Multi-Agent Transformer (MAT) to capture long-term temporal correlations among heterogeneous agents, including photovoltaics (PVs), energy storage systems, electric vehicle (EV) charging stations, and smart buildings. Simulations conducted in the PowerGridworld environment demonstrate that GMAT significantly outperforms baseline MARL algorithms, effectively suppressing voltage violations within prescribed thresholds while optimizing energy efficiency. This method provides a robust, topology-aware solution for the intelligent operation of complex distribution networks.

Keywords: Human-Centered Automation, AI-Based Methods, Building Automation
Abstract: Occupancy information has become a key component of modern building automation, as it allows environmental control systems to align their operation with actual usage patterns and optimize energy performance. WiFi channel state information (CSI) has been widely studied for device-free indoor occupancy estimation. However, effectively handling raw CSI data and extracting meaningful features remain challenging. To address this issue, we propose DBF-WiOS, a dual-branch fusion model for occupancy estimation with WiFi sensing. Specifically, amplitude and phase processing together with Short-time Fourier Transform (STFT) are applied to the raw CSI data to capture both spatial–temporal and time–frequency characteristics. A dual-branch learning architecture is then employed, consisting of a raw branch and a time–frequency branch enhanced with an attention mechanism, enabling effective feature extraction for accurate occupancy estimation. Experimental results in typical real-world indoor environments show that DBF-WiOS attains 84.5% accuracy and an F1-score of 87.7%, highlighting its capability for WiFi-based occupancy estimation.

Keywords: AI-Based Methods, Diagnosis and Prognostics, Intelligent and Flexible Manufacturing
Abstract: Milling chatter is a lasting issue that significantly affects surface quality, tool life, and machine health. However, chatter prognosis, specifically the prediction of chatter occurrence and resulting surface roughness, is challenging using conventional mechanistic and simulation methods due to complex process mechanics and strong nonlinear interactions in the cutting process. This paper develops a generative AI model to overcome the challenge of imbalanced training datasets caused by limited chatter data. Synthetic multimodal data including surface images, surface roughness, and acoustic time-series data, are generated to represent various chatter severity levels. By balancing experimental data with augmented data, the proposed multimodal generative AI model effectively predicts diverse chatter scenarios and surface roughness. For the minority chatter class, M³DCGAN augmentation improved prediction accuracy from R² = 0.4359 to 0.9667 in the small dataset and from -0.2072 to 0.9968 in the large dataset, demonstrating its effectiveness for severe data imbalance.

Keywords: Zero-Defect Manufacturing, Product Design, Development and Prototyping
Abstract: Functional ANOVA (FANOVA) is a widely used variance-based sensitivity analysis tool. However, studies on functional-output FANOVA remain relatively scarce, especially for black-box computer experiments, which often involve complex and nonlinear functional-output relationships with unknown data distribution. Conventional approaches often rely on predefined basis functions or parametric structures that lack the flexibility to capture complex nonlinear relationships. Additionally, strong assumptions about the underlying data distributions further limit their ability to achieve a data-driven orthogonal effect decomposition. To address these challenges, this study proposes a functional-output orthogonal additive Gaussian process (FOAGP) to efficiently perform the data-driven orthogonal effect decomposition. By enforcing a conditional orthogonality constraint on the separable prior process, the proposed functional-output orthogonal additive kernel enables data-driven orthogonality without requiring prior distributional assumptions. The FOAGP framework also provides analytical formulations for local Sobol' indices and expected conditional variance sensitivity indices, enabling comprehensive sensitivity analysis by capturing both global and local effect significance. Validation through two simulation studies and a real case study on fuselage shape control confirms the model's effectiveness in orthogonal effect decomposition and variance decomposition, demonstrating its practical value in engineering applications.

Keywords: Zero-Defect Manufacturing, Product Design, Development and Prototyping
Abstract: Automobile safety recalls represent critical regula tory interventions aimed at mitigating risks caused by manufacturing, design, or software defects. Accurate prediction of the time and event type for the occurrence of spatio-temporal events can be helpful in the application like disaster response and social media analysis. A novel time-aware sequential model is proposed to capture the dynamic relationship between complaint clusters and subsequent recall actions. The model leverages the Hawkes process to explicitly capture the detailed influence of historical customers’ complaints on an event-by-event basis. This model focuses on learning the dependence among events using self-attention mechanism. The fully attention-based intensity function circumvents the limitations of the parameterization and ensures that the model can effectively capture intricate intensity patterns at nonevent positions, thus enhancing the model’s expressive power.

Keywords: Diagnosis and Prognostics, Failure Detection and Recovery
Abstract: Unsupervised anomaly detection is important for industrial robot monitoring, but abnormal samples are rare, difficult to collect, and vary widely across fault types in real production. Existing data-driven methods often operate as black-box models and make limited use of robot physical operating knowledge. We proposed a hybrid dynamics-aware framework for unsupervised indutrial robot anomaly detection. The proposed approach integrates physical prior information with temporal modeling to improve sensitivity to subtle abnormal deviations. Evaluation on a robot dataset shows that the proposed method outperforms the strongest reported baseline and achieves more stable performance across anomaly categories. The framework provides a promising approach for practical industrial robot health monitoring.

Keywords: Big-Data and Data Mining, AI-Based Methods, Product Design, Development and Prototyping
Abstract: In electrical drive bench testing, a large volume of measurement data and unstructured document data are being generated, but are still underutilized in data intelligence. With continuous iterative development, Retrieval-Augmented Generation (RAG) based on Large Language Models provides a powerful automated data integration solution for data with multiple modalities. However, the particular irregularity and heterogeneity of e-drive bench test data poses challenges to the practical implementation of an effective RAG solution in this specific domain and requires practice-oriented adaptations of RAG components and strategies. To address these challenges, the role of metadata in advanced retrieval strategies is investigated. A set-theoretic framework for structuring metadata in this industry vertical domain is proposed. Classical RAG components that incorporate extracted metadata are evaluated to identify the optimal approach for practical e-drive bench test data retrieval scenarios.

Keywords: Control Architectures and Programming, Sensor-based Control, Intelligent Transportation Systems
Abstract: Magnetic levitation (maglev) cars integrate levitation, propulsion, and guidance in a unified system and are regarded as a promising solution for future intelligent transportation. Although extensive studies have been conducted on intelligent control strategies for maglev systems, most existing work focuses on levitation stability rather than longitudinal speed regulation. During vehicle operation, however, precise speed tracking is critical to safety, especially under nonlinear coupling effects in driving and braking processes. To address overshoot, chattering, and steady-state error caused by nonlinear dynamics, this paper proposes a composite speed tracking control strategy that integrates a tracking differentiator (TD) with fractional-order integral non-singular terminal sliding mode control (FOI-NTSMC). A data-driven nonlinear longitudinal dynamic model is established using permanent magnet electrodynamic wheel (PMEDW) driving force experimental data and the least squares method. The TD preprocesses the reference speed signal, while a fractional-order integral sliding mode surface and reaching law are designed to enhance robustness. A saturation function is introduced to suppress chattering and compensate for nonlinear coupling disturbances. Simulation studies indicate, compared with PID control, traditional sliding mode control, and non-singular terminal sliding mode control, demonstrate overshoot-free tracking, zero steady-state error, and a 15%–30% improvement in dynamic response over conventional NTSMC. The proposed strategy provides an effective solution for high-precision longitudinal speed control and fast braking of maglev cars.

Keywords: Intelligent Transportation Systems, Power and Energy Systems automation, Optimization and Optimal Control
Abstract: High-Speed Railway (HSR) Traction Power Supply Systems (TPSS) face unprecedented degradation risks under the coupled impacts of extreme weather (e.g., typhoons) and heavy structural transportation loads. Traditional singular resilience assessments frequently underestimate these compounding vulnerabilities. To address this gap, this paper develops a dynamic, dual-stressor risk index model that mathematically couples aerodynamic stresses with severe electro-thermal fatigue. The degradation process is evaluated through a multi-state Sequential Monte Carlo (SMC) simulation that specifically accounts for weather-restricted repair constraints. To prevent non-linear cascading system collapse under these coupled shocks, a novel ``Hardware-Software'' coordinated resilience enhancement paradigm is proposed. This approach distinctively integrates physical localized Emergency Energy Management (EEM) with active algorithmic Adaptive Train Control (ATC). Utilizing an objective Entropy Weight-TOPSIS multi-criteria decision-making framework, an ablation study compares four resilience strategies across three critical operational dimensions: resilience integral, extreme survivability, and severely degraded time. Simulation results quantitatively demonstrate that isolated hardware or software solutions are insufficient to withstand complex multi-stressor surges. Alternatively, the proposed coordinated (EEM+ATC) strategy successfully linearizes system degradation, prevents widespread outages, mitigates operational risks associated with passenger delays by notably curtailing severely degraded time from 11.27h to 4.65h, and comprehensively maximizes operational resilience during extreme disaster events.

Keywords: Intelligent Transportation Systems, Hybrid Logical/Dynamical Planning and Verification, Collision Avoidance
Abstract: Virtual coupling (VC) enables trains to operate with reduced headways, thereby improving line capacity and operational flexibility. However, the close coordination among multiple trains and the switching of operating modes significantly increase operational risks. To address safety issues in virtual coupling operations, this paper proposes a real-time train safety verification method based on hybrid automata reachability analysis. The proposed approach models train behaviors under different operation modes with a hybrid automaton that integrates discrete operating modes with continuoustime longitudinal train dynamics. By computing the reachable set of the model over a given future time interval and checking its intersection with predefined unsafe sets, potential risks of overspeed and inter-train collisions can be identified in real time. Simulation results demonstrate that the proposed method effectively detects hazardous train behaviors while satisfying the computational constraints of real-time verification.

Keywords: Behavior-Based Systems, Discrete Event Dynamic Automation Systems, Reinforcement
Abstract: The proliferation of mobile devices has made location-based services an integral part of daily life, yet continuous sharing of precise locations poses significant privacy risks, such as exposure to unwarranted surveillance. To address this challenge, this paper proposes a novel reinforcement learning- based method to generate an optimal insertion function under a private insertion mechanism, ensuring current-state opacity without requiring a formal model of user mobility. Using the publicly available Geolife GPS trajectory dataset, the density-based spatial clustering of applications with noise algorithm, parameterized via a K-distance graph, is applied to extract meaningful location states from raw trajectories. Experimental results show that as training proceeds, the cumulative reward converges and the probability of secret exposure approaches zero. Furthermore, the resulting insertion automaton is isomorphic to that obtained by the baseline model-based method, demonstrating the effectiveness of our data-driven approach.

Keywords: Logistics, Planning, Scheduling and Coordination, Intelligent Transportation Systems
Abstract: This paper introduces a collaborative mode between ground rescue vehicle (GRV) and unmanned aerial vehicle (UAV) into emergency delivery, aiming to alleviate the action burden on rescue personnel through collaborative transportation. In the constructed collaborative model, multiple UAVs and GRVs are included, with UAVs capable of visiting multiple task nodes after each launch. The GRV is not required to complete the UAV's support operations at the same node, and the collaborative process is dynamic. The GRV can synchronously perform rescue tasks, while the UAV can also visit some ineligible nodes to obtain support. These considerations are expected to enhance the flexibility of the collaborative process. To solve this model, a hybrid heuristic algorithm with a synchronization strategy is designed, comprising three components: route planning, time synchronization, and heuristic search. Compared to multiple benchmark algorithms, case test results verify the effectiveness of the proposed method. Through testing multiple parameter combinations and real-world cases, detailed analysis of collaboration characteristics under varying capabilities was conducted, providing management insights for applying collaboration modes in emergency delivery scenarios.